diff --git a/document.aux b/document.aux index 713dc6d..a7544e0 100644 --- a/document.aux +++ b/document.aux @@ -143,9 +143,9 @@ \abx@aux@segm{0}{0}{29} \abx@aux@cite{30} \abx@aux@segm{0}{0}{30} -\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsubsection}{Optimisers}{20}{section*.25}} \abx@aux@cite{31} \abx@aux@segm{0}{0}{31} +\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsubsection}{Optimisers}{20}{section*.25}} \abx@aux@segm{0}{0}{30} \abx@aux@cite{32} \abx@aux@segm{0}{0}{32} @@ -164,7 +164,6 @@ \@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsection}{Data pre-processing}{24}{section*.30}} \@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsection}{Spam Filtering}{24}{section*.31}} \@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsection}{Language Detection}{24}{section*.32}} -\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsection}{Sentiment Analysis}{24}{section*.33}} \abx@aux@segm{0}{0}{12} \abx@aux@segm{0}{0}{11} \abx@aux@segm{0}{0}{22} @@ -172,48 +171,51 @@ \abx@aux@segm{0}{0}{25} \abx@aux@segm{0}{0}{22} \abx@aux@segm{0}{0}{25} +\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsection}{Sentiment Analysis}{25}{section*.33}} \@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsection}{Neural Network}{25}{section*.34}} -\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsection}{Price Forecasting}{26}{section*.35}} -\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsection}{Frontend Application}{26}{section*.36}} -\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsection}{With reference to Initial PID}{26}{section*.37}} -\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsection}{Solution Summary}{27}{section*.38}} -\newlabel{summary}{{}{27}{Solution Summary}{section*.38}{}} -\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsection}{Data flow Overview}{27}{section*.39}} -\newlabel{data-flow}{{}{27}{Data flow Overview}{section*.39}{}} -\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {section}{System Design}{28}{section*.40}} -\newlabel{Design}{{}{28}{System Design}{section*.40}{}} -\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsection}{Dataflow Designs}{28}{section*.41}} +\abx@aux@cite{34} +\abx@aux@segm{0}{0}{34} +\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsection}{Price Forecasting}{26}{section*.36}} +\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsection}{Frontend Application}{26}{section*.37}} +\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsection}{With reference to Initial PID}{27}{section*.38}} +\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsection}{Solution Summary}{27}{section*.39}} +\newlabel{summary}{{}{27}{Solution Summary}{section*.39}{}} +\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsection}{Data flow Overview}{27}{section*.40}} +\newlabel{data-flow}{{}{27}{Data flow Overview}{section*.40}{}} +\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {section}{System Design}{29}{section*.41}} +\newlabel{Design}{{}{29}{System Design}{section*.41}{}} +\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsection}{Dataflow Designs}{29}{section*.42}} \abx@aux@segm{0}{0}{12} -\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsection}{UML Component Design}{34}{section*.42}} -\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsection}{Interface Design}{34}{section*.43}} -\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {section}{Implementation}{35}{section*.44}} -\newlabel{implementation}{{}{35}{Implementation}{section*.44}{}} -\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsection}{Data collection}{35}{section*.45}} -\newlabel{collection}{{}{35}{Data collection}{section*.45}{}} -\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsubsection}{Price Time-series Data}{35}{section*.46}} -\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsection}{Data processing}{35}{section*.47}} -\newlabel{processing}{{}{35}{Data processing}{section*.47}{}} -\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsubsection}{Preprocessing}{35}{section*.48}} -\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {paragraph}{Tweet Filtering}{35}{section*.49}} -\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {paragraph}{Text Cleaning}{35}{section*.50}} -\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {paragraph}{Ngram based Language detection filtering}{35}{section*.51}} -\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsubsection}{Spam Filtering}{35}{section*.52}} -\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {paragraph}{Tweet Processing}{35}{section*.53}} -\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {paragraph}{Naive Bayes model}{35}{section*.54}} -\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsection}{Sentiment Analysis}{35}{section*.55}} -\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsubsection}{VADER}{35}{section*.56}} -\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsection}{Recurrent Neural Network - LSTM}{35}{section*.57}} -\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsubsection}{Training and Testing Model}{35}{section*.58}} -\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsubsection}{Scoring and Validation}{35}{section*.59}} -\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsubsection}{Future Prediction Forecasting}{35}{section*.60}} -\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {section}{Testing: Verification and Reflection}{36}{section*.61}} -\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {section}{Discussion: Contribution and Reflection}{37}{section*.62}} -\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsection}{Limitations}{37}{section*.63}} -\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {section}{Conclusion and Future Improvements}{38}{section*.64}} -\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsection}{Conclusion}{38}{section*.65}} -\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsection}{Future Improvements}{38}{section*.66}} -\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {section}{Appendices}{42}{section*.68}} -\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsection}{Appendix A - Project Initiation Document}{42}{section*.69}} +\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsection}{UML Component Design}{35}{section*.43}} +\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsection}{Interface Design}{35}{section*.44}} +\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {section}{Implementation}{36}{section*.46}} +\newlabel{implementation}{{}{36}{Implementation}{section*.46}{}} +\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsection}{Data collection}{36}{section*.47}} +\newlabel{collection}{{}{36}{Data collection}{section*.47}{}} +\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsubsection}{Price Time-series Data}{36}{section*.48}} +\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsection}{Data processing}{36}{section*.49}} +\newlabel{processing}{{}{36}{Data processing}{section*.49}{}} +\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsubsection}{Preprocessing}{36}{section*.50}} +\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {paragraph}{Tweet Filtering}{36}{section*.51}} +\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {paragraph}{Text Cleaning}{36}{section*.52}} +\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {paragraph}{Ngram based Language detection filtering}{36}{section*.53}} +\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsubsection}{Spam Filtering}{36}{section*.54}} +\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {paragraph}{Tweet Processing}{36}{section*.55}} +\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {paragraph}{Naive Bayes model}{36}{section*.56}} +\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsection}{Sentiment Analysis}{36}{section*.57}} +\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsubsection}{VADER}{36}{section*.58}} +\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsection}{Recurrent Neural Network - LSTM}{36}{section*.59}} +\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsubsection}{Training and Testing Model}{36}{section*.60}} +\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsubsection}{Scoring and Validation}{36}{section*.61}} +\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsubsection}{Future Prediction Forecasting}{36}{section*.62}} +\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {section}{Testing: Verification and Reflection}{37}{section*.63}} +\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {section}{Discussion: Contribution and Reflection}{38}{section*.64}} +\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsection}{Limitations}{38}{section*.65}} +\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {section}{Conclusion and Future Improvements}{39}{section*.66}} +\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsection}{Conclusion}{39}{section*.67}} +\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsection}{Future Improvements}{39}{section*.68}} +\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {section}{Appendices}{43}{section*.70}} +\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsection}{Appendix A - Project Initiation Document}{43}{section*.71}} \abx@aux@refcontextdefaultsdone \abx@aux@defaultrefcontext{0}{1}{none/global//global/global} \abx@aux@defaultrefcontext{0}{2}{none/global//global/global} @@ -248,4 +250,5 @@ \abx@aux@defaultrefcontext{0}{31}{none/global//global/global} \abx@aux@defaultrefcontext{0}{32}{none/global//global/global} \abx@aux@defaultrefcontext{0}{33}{none/global//global/global} -\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsection}{Appendix B - Log book}{55}{section*.70}} +\abx@aux@defaultrefcontext{0}{34}{none/global//global/global} +\@writefile{toc}{\defcounter {refsection}{0}\relax }\@writefile{toc}{\contentsline {subsection}{Appendix B - Log book}{56}{section*.72}} diff --git a/document.bbl b/document.bbl index 219c395..fdde499 100644 --- a/document.bbl +++ b/document.bbl @@ -1126,6 +1126,36 @@ \verb http://docs.tweepy.org/en/v3.5.0/ \endverb \endentry + \entry{34}{inproceedings}{} + \name{author}{1}{}{% + {{hash=974f50284a1994b00c1f04d211402eb0}{% + family={Deoras}, + familyi={D\bibinitperiod}, + given={Srishti}, + giveni={S\bibinitperiod}}}% + } + \list{organization}{1}{% + {Analytics India}% + } + \strng{namehash}{974f50284a1994b00c1f04d211402eb0} + \strng{fullhash}{974f50284a1994b00c1f04d211402eb0} + \strng{bibnamehash}{974f50284a1994b00c1f04d211402eb0} + \strng{authorbibnamehash}{974f50284a1994b00c1f04d211402eb0} + \strng{authornamehash}{974f50284a1994b00c1f04d211402eb0} + \strng{authorfullhash}{974f50284a1994b00c1f04d211402eb0} + \field{sortinit}{6} + \field{sortinithash}{57e57fb8451e7fcfa45d1e069f6d3136} + \field{labelnamesource}{author} + \field{labeltitlesource}{title} + \field{title}{Tensorflow Vs. Theano: What Do Researchers Prefer As An Artificial Intelligence Framework} + \field{year}{2017} + \verb{urlraw} + \verb https://www.analyticsindiamag.com/tensorflow-vs-theano-researchers-prefer-artificial-intelligence-framework + \endverb + \verb{url} + \verb https://www.analyticsindiamag.com/tensorflow-vs-theano-researchers-prefer-artificial-intelligence-framework + \endverb + \endentry \enddatalist \endrefsection \endinput diff --git a/document.bcf b/document.bcf index f765db0..5df50bb 100644 --- a/document.bcf +++ b/document.bcf @@ -1990,7 +1990,8 @@ 25 22 25 - 12 + 34 + 12 * diff --git a/document.blg b/document.blg index d14ee32..815288a 100644 --- a/document.blg +++ b/document.blg @@ -1,20 +1,20 @@ [0] Config.pm:302> INFO - This is Biber 2.9 [0] Config.pm:305> INFO - Logfile is 'document.blg' -[20] biber:313> INFO - === Wed Apr 24, 2019, 16:48:01 -[35] Biber.pm:371> INFO - Reading 'document.bcf' -[84] Biber.pm:854> INFO - Using all citekeys in bib section 0 -[95] Biber.pm:3981> INFO - Processing section 0 -[103] Biber.pm:4154> INFO - Looking for bibtex format file 'report.bib' for section 0 -[105] bibtex.pm:1468> INFO - LaTeX decoding ... -[134] bibtex.pm:1294> INFO - Found BibTeX data source 'report.bib' -[184] Utils.pm:169> WARN - Name "Mairal, J., Ponce, J., Sapiro, G., Zisserman, A." has too many commas: skipping name -[235] Utils.pm:169> WARN - year field 'Mar 13, 2016' in entry '23' is not an integer - this will probably not sort properly. -[249] Utils.pm:169> WARN - BibTeX subsystem: warning: comma(s) at end of name (removing) -[249] Utils.pm:169> WARN - BibTeX subsystem: author, warning: comma(s) at end of name (removing) -[268] UCollate.pm:68> INFO - Overriding locale 'en-US' defaults 'normalization = NFD' with 'normalization = prenormalized' -[268] UCollate.pm:68> INFO - Overriding locale 'en-US' defaults 'variable = shifted' with 'variable = non-ignorable' -[268] Biber.pm:3809> INFO - Sorting list 'none/global//global/global' of type 'entry' with template 'none' and locale 'en-US' -[268] Biber.pm:3815> INFO - No sort tailoring available for locale 'en-US' -[288] bbl.pm:617> INFO - Writing 'document.bbl' with encoding 'ascii' -[299] bbl.pm:720> INFO - Output to document.bbl -[300] Biber.pm:110> INFO - WARNINGS: 4 +[18] biber:313> INFO - === Thu Apr 25, 2019, 00:03:40 +[33] Biber.pm:371> INFO - Reading 'document.bcf' +[79] Biber.pm:854> INFO - Using all citekeys in bib section 0 +[88] Biber.pm:3981> INFO - Processing section 0 +[96] Biber.pm:4154> INFO - Looking for bibtex format file 'report.bib' for section 0 +[97] bibtex.pm:1468> INFO - LaTeX decoding ... +[125] bibtex.pm:1294> INFO - Found BibTeX data source 'report.bib' +[133] Utils.pm:169> WARN - year field 'Mar 13, 2016' in entry '23' is not an integer - this will probably not sort properly. +[206] Utils.pm:169> WARN - Name "Mairal, J., Ponce, J., Sapiro, G., Zisserman, A." has too many commas: skipping name +[238] Utils.pm:169> WARN - BibTeX subsystem: warning: comma(s) at end of name (removing) +[238] Utils.pm:169> WARN - BibTeX subsystem: author, warning: comma(s) at end of name (removing) +[257] UCollate.pm:68> INFO - Overriding locale 'en-US' defaults 'variable = shifted' with 'variable = non-ignorable' +[257] UCollate.pm:68> INFO - Overriding locale 'en-US' defaults 'normalization = NFD' with 'normalization = prenormalized' +[257] Biber.pm:3809> INFO - Sorting list 'none/global//global/global' of type 'entry' with template 'none' and locale 'en-US' +[257] Biber.pm:3815> INFO - No sort tailoring available for locale 'en-US' +[275] bbl.pm:617> INFO - Writing 'document.bbl' with encoding 'ascii' +[287] bbl.pm:720> INFO - Output to document.bbl +[288] Biber.pm:110> INFO - WARNINGS: 4 diff --git a/document.log b/document.log index 0faf0e2..4607693 100644 --- a/document.log +++ b/document.log @@ -1,4 +1,4 @@ -This is pdfTeX, Version 3.14159265-2.6-1.40.18 (TeX Live 2017/Debian) (preloaded format=pdflatex 2018.10.16) 24 APR 2019 19:04 +This is pdfTeX, Version 3.14159265-2.6-1.40.18 (TeX Live 2017/Debian) (preloaded format=pdflatex 2018.10.16) 25 APR 2019 00:06 entering extended mode restricted \write18 enabled. %&-line parsing enabled. @@ -1005,16 +1005,13 @@ Underfull \hbox (badness 10000) in paragraph at lines 216--218 [] [12] [13] [14] [15] -Missing character: There is no â in font cmr10! -Missing character: There is no € in font cmr10! -Missing character: There is no ™ in font cmr10! - [16] - + File: images/perceptron.png Graphic file (type png) Package pdftex.def Info: images/perceptron.png used on input line 308. (pdftex.def) Requested size: 284.52713pt x 170.72142pt. - + [16] + File: images/rnn_ffn.png Graphic file (type png) Package pdftex.def Info: images/rnn_ffn.png used on input line 328. @@ -1029,11 +1026,10 @@ Package pdftex.def Info: images/lstm.png used on input line 346. Missing character: There is no â in font cmr10! Missing character: There is no € in font cmr10! Missing character: There is no ™ in font cmr10! - [20] Missing character: There is no â in font cmr10! Missing character: There is no € in font cmr10! Missing character: There is no ™ in font cmr10! - [21] [22] + [20] [21] [22] Underfull \hbox (badness 10000) in paragraph at lines 456--458 [] @@ -1061,106 +1057,117 @@ Underfull \hbox (badness 10000) in paragraph at lines 471--475 [23] LaTeX Font Info: Font shape `OMS/cmr/m/it' in size <10> not available (Font) Font shape `OMS/cmsy/m/n' tried instead on input line 488. - [24] [25] [26] - + [24] [25] +Missing character: There is no â in font cmr10! +Missing character: There is no € in font cmr10! +Missing character: There is no ™ in font cmr10! + [26] + File: images/Generic_Flow.png Graphic file (type png) -Package pdftex.def Info: images/Generic_Flow.png used on input line 565. +Package pdftex.def Info: images/Generic_Flow.png used on input line 574. (pdftex.def) Requested size: 512.13474pt x 227.62009pt. -Overfull \hbox (71.28728pt too wide) in paragraph at lines 565--566 +Overfull \hbox (71.28728pt too wide) in paragraph at lines 574--575 [][] [] -[27 <./images/Generic_Flow.png (PNG copy)>] -Underfull \hbox (badness 10000) in paragraph at lines 576--578 +[27] [28 <./images/Generic_Flow.png (PNG copy)>] +Underfull \hbox (badness 10000) in paragraph at lines 585--587 [] - + File: images/Dataflow.png Graphic file (type png) -Package pdftex.def Info: images/Dataflow.png used on input line 581. +Package pdftex.def Info: images/Dataflow.png used on input line 590. (pdftex.def) Requested size: 512.09683pt x 227.62125pt. -Overfull \hbox (71.24937pt too wide) in paragraph at lines 581--583 +Overfull \hbox (71.24937pt too wide) in paragraph at lines 590--592 [] [] -[28 <./images/Dataflow.png (PNG copy)>] - +[29 <./images/Dataflow.png (PNG copy)>] + File: images/Data_Collector.png Graphic file (type png) -Package pdftex.def Info: images/Data_Collector.png used on input line 589. +Package pdftex.def Info: images/Data_Collector.png used on input line 598. (pdftex.def) Requested size: 426.78574pt x 227.61746pt. - [29 <./images/Data_Collector.png (PNG copy)>] - + [30 <./images/Data_Collector.png (PNG copy)>] + File: images/Analysis_Engine.png Graphic file (type png) -Package pdftex.def Info: images/Analysis_Engine.png used on input line 604. +Package pdftex.def Info: images/Analysis_Engine.png used on input line 613. (pdftex.def) Requested size: 483.67276pt x 227.62561pt. -Overfull \hbox (42.8253pt too wide) in paragraph at lines 604--606 +Overfull \hbox (42.8253pt too wide) in paragraph at lines 613--615 [] [] -[30 <./images/Analysis_Engine.png (PNG copy)>] - +[31 <./images/Analysis_Engine.png (PNG copy)>] + File: images/Neural_Network.png Graphic file (type png) -Package pdftex.def Info: images/Neural_Network.png used on input line 620. +Package pdftex.def Info: images/Neural_Network.png used on input line 629. (pdftex.def) Requested size: 483.6893pt x 341.42757pt. -Overfull \hbox (42.84184pt too wide) in paragraph at lines 620--622 +Overfull \hbox (42.84184pt too wide) in paragraph at lines 629--631 [] [] -[31 <./images/Neural_Network.png (PNG copy)>] - +[32 <./images/Neural_Network.png (PNG copy)>] + File: images/Future_Predictions.png Graphic file (type png) -Package pdftex.def Info: images/Future_Predictions.png used on input line 634. +Package pdftex.def Info: images/Future_Predictions.png used on input line 643. (pdftex.def) Requested size: 512.1362pt x 227.62119pt. -Overfull \hbox (71.28874pt too wide) in paragraph at lines 634--636 +Overfull \hbox (71.28874pt too wide) in paragraph at lines 643--645 [] [] -[32 <./images/Future_Predictions.png (PNG copy)>] - +[33 <./images/Future_Predictions.png (PNG copy)>] + File: images/Frontend_Application.png Graphic file (type png) -Package pdftex.def Info: images/Frontend_Application.png used on input line 64 -8. +Package pdftex.def Info: images/Frontend_Application.png used on input line 65 +7. (pdftex.def) Requested size: 284.52162pt x 256.07664pt. - [33 <./images/Frontend_Application.png (PNG copy)>] [34] [35] [36] [37] [38] -Overfull \hbox (1.46323pt too wide) in paragraph at lines 717--717 + [34 <./images/Frontend_Application.png (PNG copy)>] + +File: images/interface_design.png Graphic file (type png) + +Package pdftex.def Info: images/interface_design.png used on input line 676. +(pdftex.def) Requested size: 227.61479pt x 369.88063pt. + [35 <./images/interface_design.png>] [36] [37] [38] +[39] +Overfull \hbox (1.46323pt too wide) in paragraph at lines 736--736 \OT1/cmr/m/n/10 [On-line]. Avail-able: []$\OT1/cmtt/m/n/10 https : / / www . co deproject . com / Articles / 1201444 / Stock -[] Predictions -[] [] -[39] -Overfull \hbox (22.26572pt too wide) in paragraph at lines 717--717 +[40] +Overfull \hbox (22.26572pt too wide) in paragraph at lines 736--736 \OT1/cmr/m/n/10 able: []$\OT1/cmtt/m/n/10 https : / / towardsdatascience . com / recurrent -[] neural -[] networks -[] and -[] lstm -[] 4b601dd822a5$[]\OT1/cm r/m/n/10 . [] -Overfull \hbox (59.98181pt too wide) in paragraph at lines 717--717 +Overfull \hbox (59.98181pt too wide) in paragraph at lines 736--736 []$\OT1/cmtt/m/n/10 http : / / deeplearning . stanford . edu / tutorial / super vised / OptimizationStochasticGradientDescent$[]\OT1/cmr/m/n/10 . [] -[40] [41] +[41] [42] pdfTeX warning: /usr/bin/pdflatex (file ./PID.pdf): PDF inclusion: found PDF ve rsion <1.7>, but at most version <1.5> allowed - + File: PID.pdf Graphic file (type pdf) -Package pdftex.def Info: PID.pdf used on input line 722. +Package pdftex.def Info: PID.pdf used on input line 741. (pdftex.def) Requested size: 597.551pt x 845.07512pt. @@ -1168,7 +1175,7 @@ pdfTeX warning: /usr/bin/pdflatex (file ./PID.pdf): PDF inclusion: found PDF ve rsion <1.7>, but at most version <1.5> allowed File: PID.pdf Graphic file (type pdf) -Package pdftex.def Info: PID.pdf used on input line 722. +Package pdftex.def Info: PID.pdf used on input line 741. (pdftex.def) Requested size: 597.551pt x 845.07512pt. @@ -1178,233 +1185,233 @@ rsion <1.7>, but at most version <1.5> allowed pdfTeX warning: /usr/bin/pdflatex (file ./PID.pdf): PDF inclusion: found PDF ve rsion <1.7>, but at most version <1.5> allowed - + File: PID.pdf Graphic file (type pdf) -Package pdftex.def Info: PID.pdf , page1 used on input line 722. +Package pdftex.def Info: PID.pdf , page1 used on input line 741. (pdftex.def) Requested size: 597.551pt x 845.07512pt. File: PID.pdf Graphic file (type pdf) -Package pdftex.def Info: PID.pdf , page1 used on input line 722. +Package pdftex.def Info: PID.pdf , page1 used on input line 741. (pdftex.def) Requested size: 562.1644pt x 795.0303pt. -[42] +[43] File: PID.pdf Graphic file (type pdf) -Package pdftex.def Info: PID.pdf , page1 used on input line 722. +Package pdftex.def Info: PID.pdf , page1 used on input line 741. (pdftex.def) Requested size: 562.1644pt x 795.0303pt. File: PID.pdf Graphic file (type pdf) -Package pdftex.def Info: PID.pdf , page1 used on input line 722. +Package pdftex.def Info: PID.pdf , page1 used on input line 741. (pdftex.def) Requested size: 562.1644pt x 795.0303pt. File: PID.pdf Graphic file (type pdf) -Package pdftex.def Info: PID.pdf , page1 used on input line 722. +Package pdftex.def Info: PID.pdf , page1 used on input line 741. (pdftex.def) Requested size: 562.1644pt x 795.0303pt. - [43 <./PID.pdf>] + [44 <./PID.pdf>] pdfTeX warning: /usr/bin/pdflatex (file ./PID.pdf): PDF inclusion: found PDF ve rsion <1.7>, but at most version <1.5> allowed - + File: PID.pdf Graphic file (type pdf) -Package pdftex.def Info: PID.pdf , page2 used on input line 722. +Package pdftex.def Info: PID.pdf , page2 used on input line 741. (pdftex.def) Requested size: 562.1644pt x 795.0303pt. File: PID.pdf Graphic file (type pdf) -Package pdftex.def Info: PID.pdf , page2 used on input line 722. +Package pdftex.def Info: PID.pdf , page2 used on input line 741. (pdftex.def) Requested size: 562.1644pt x 795.0303pt. File: PID.pdf Graphic file (type pdf) -Package pdftex.def Info: PID.pdf , page2 used on input line 722. -(pdftex.def) Requested size: 562.1644pt x 795.0303pt. -[44 <./PID.pdf>] - -pdfTeX warning: /usr/bin/pdflatex (file ./PID.pdf): PDF inclusion: found PDF ve -rsion <1.7>, but at most version <1.5> allowed - -File: PID.pdf Graphic file (type pdf) - -Package pdftex.def Info: PID.pdf , page3 used on input line 722. -(pdftex.def) Requested size: 562.1644pt x 795.0303pt. -File: PID.pdf Graphic file (type pdf) - -Package pdftex.def Info: PID.pdf , page3 used on input line 722. -(pdftex.def) Requested size: 562.1644pt x 795.0303pt. -File: PID.pdf Graphic file (type pdf) - -Package pdftex.def Info: PID.pdf , page3 used on input line 722. +Package pdftex.def Info: PID.pdf , page2 used on input line 741. (pdftex.def) Requested size: 562.1644pt x 795.0303pt. [45 <./PID.pdf>] pdfTeX warning: /usr/bin/pdflatex (file ./PID.pdf): PDF inclusion: found PDF ve rsion <1.7>, but at most version <1.5> allowed - + File: PID.pdf Graphic file (type pdf) - -Package pdftex.def Info: PID.pdf , page4 used on input line 722. + +Package pdftex.def Info: PID.pdf , page3 used on input line 741. (pdftex.def) Requested size: 562.1644pt x 795.0303pt. File: PID.pdf Graphic file (type pdf) - -Package pdftex.def Info: PID.pdf , page4 used on input line 722. + +Package pdftex.def Info: PID.pdf , page3 used on input line 741. (pdftex.def) Requested size: 562.1644pt x 795.0303pt. File: PID.pdf Graphic file (type pdf) - -Package pdftex.def Info: PID.pdf , page4 used on input line 722. + +Package pdftex.def Info: PID.pdf , page3 used on input line 741. (pdftex.def) Requested size: 562.1644pt x 795.0303pt. [46 <./PID.pdf>] pdfTeX warning: /usr/bin/pdflatex (file ./PID.pdf): PDF inclusion: found PDF ve rsion <1.7>, but at most version <1.5> allowed - + File: PID.pdf Graphic file (type pdf) - -Package pdftex.def Info: PID.pdf , page5 used on input line 722. + +Package pdftex.def Info: PID.pdf , page4 used on input line 741. (pdftex.def) Requested size: 562.1644pt x 795.0303pt. File: PID.pdf Graphic file (type pdf) - -Package pdftex.def Info: PID.pdf , page5 used on input line 722. + +Package pdftex.def Info: PID.pdf , page4 used on input line 741. (pdftex.def) Requested size: 562.1644pt x 795.0303pt. File: PID.pdf Graphic file (type pdf) - -Package pdftex.def Info: PID.pdf , page5 used on input line 722. + +Package pdftex.def Info: PID.pdf , page4 used on input line 741. (pdftex.def) Requested size: 562.1644pt x 795.0303pt. [47 <./PID.pdf>] pdfTeX warning: /usr/bin/pdflatex (file ./PID.pdf): PDF inclusion: found PDF ve rsion <1.7>, but at most version <1.5> allowed - + File: PID.pdf Graphic file (type pdf) - -Package pdftex.def Info: PID.pdf , page6 used on input line 722. + +Package pdftex.def Info: PID.pdf , page5 used on input line 741. (pdftex.def) Requested size: 562.1644pt x 795.0303pt. File: PID.pdf Graphic file (type pdf) - -Package pdftex.def Info: PID.pdf , page6 used on input line 722. + +Package pdftex.def Info: PID.pdf , page5 used on input line 741. (pdftex.def) Requested size: 562.1644pt x 795.0303pt. File: PID.pdf Graphic file (type pdf) - -Package pdftex.def Info: PID.pdf , page6 used on input line 722. + +Package pdftex.def Info: PID.pdf , page5 used on input line 741. (pdftex.def) Requested size: 562.1644pt x 795.0303pt. [48 <./PID.pdf>] pdfTeX warning: /usr/bin/pdflatex (file ./PID.pdf): PDF inclusion: found PDF ve rsion <1.7>, but at most version <1.5> allowed - + File: PID.pdf Graphic file (type pdf) - -Package pdftex.def Info: PID.pdf , page7 used on input line 722. + +Package pdftex.def Info: PID.pdf , page6 used on input line 741. (pdftex.def) Requested size: 562.1644pt x 795.0303pt. File: PID.pdf Graphic file (type pdf) - -Package pdftex.def Info: PID.pdf , page7 used on input line 722. + +Package pdftex.def Info: PID.pdf , page6 used on input line 741. (pdftex.def) Requested size: 562.1644pt x 795.0303pt. File: PID.pdf Graphic file (type pdf) - -Package pdftex.def Info: PID.pdf , page7 used on input line 722. + +Package pdftex.def Info: PID.pdf , page6 used on input line 741. (pdftex.def) Requested size: 562.1644pt x 795.0303pt. [49 <./PID.pdf>] pdfTeX warning: /usr/bin/pdflatex (file ./PID.pdf): PDF inclusion: found PDF ve rsion <1.7>, but at most version <1.5> allowed - + File: PID.pdf Graphic file (type pdf) - -Package pdftex.def Info: PID.pdf , page8 used on input line 722. -(pdftex.def) Requested size: 795.0303pt x 562.1644pt. + +Package pdftex.def Info: PID.pdf , page7 used on input line 741. +(pdftex.def) Requested size: 562.1644pt x 795.0303pt. File: PID.pdf Graphic file (type pdf) - -Package pdftex.def Info: PID.pdf , page8 used on input line 722. -(pdftex.def) Requested size: 795.0303pt x 562.1644pt. + +Package pdftex.def Info: PID.pdf , page7 used on input line 741. +(pdftex.def) Requested size: 562.1644pt x 795.0303pt. File: PID.pdf Graphic file (type pdf) - -Package pdftex.def Info: PID.pdf , page8 used on input line 722. -(pdftex.def) Requested size: 795.0303pt x 562.1644pt. + +Package pdftex.def Info: PID.pdf , page7 used on input line 741. +(pdftex.def) Requested size: 562.1644pt x 795.0303pt. [50 <./PID.pdf>] pdfTeX warning: /usr/bin/pdflatex (file ./PID.pdf): PDF inclusion: found PDF ve rsion <1.7>, but at most version <1.5> allowed - + File: PID.pdf Graphic file (type pdf) - -Package pdftex.def Info: PID.pdf , page9 used on input line 722. + +Package pdftex.def Info: PID.pdf , page8 used on input line 741. (pdftex.def) Requested size: 795.0303pt x 562.1644pt. File: PID.pdf Graphic file (type pdf) - -Package pdftex.def Info: PID.pdf , page9 used on input line 722. + +Package pdftex.def Info: PID.pdf , page8 used on input line 741. (pdftex.def) Requested size: 795.0303pt x 562.1644pt. File: PID.pdf Graphic file (type pdf) - -Package pdftex.def Info: PID.pdf , page9 used on input line 722. + +Package pdftex.def Info: PID.pdf , page8 used on input line 741. (pdftex.def) Requested size: 795.0303pt x 562.1644pt. [51 <./PID.pdf>] pdfTeX warning: /usr/bin/pdflatex (file ./PID.pdf): PDF inclusion: found PDF ve rsion <1.7>, but at most version <1.5> allowed - + File: PID.pdf Graphic file (type pdf) - -Package pdftex.def Info: PID.pdf , page10 used on input line 722. + +Package pdftex.def Info: PID.pdf , page9 used on input line 741. (pdftex.def) Requested size: 795.0303pt x 562.1644pt. File: PID.pdf Graphic file (type pdf) - -Package pdftex.def Info: PID.pdf , page10 used on input line 722. + +Package pdftex.def Info: PID.pdf , page9 used on input line 741. (pdftex.def) Requested size: 795.0303pt x 562.1644pt. File: PID.pdf Graphic file (type pdf) - -Package pdftex.def Info: PID.pdf , page10 used on input line 722. + +Package pdftex.def Info: PID.pdf , page9 used on input line 741. (pdftex.def) Requested size: 795.0303pt x 562.1644pt. [52 <./PID.pdf>] pdfTeX warning: /usr/bin/pdflatex (file ./PID.pdf): PDF inclusion: found PDF ve rsion <1.7>, but at most version <1.5> allowed - + File: PID.pdf Graphic file (type pdf) - -Package pdftex.def Info: PID.pdf , page11 used on input line 722. + +Package pdftex.def Info: PID.pdf , page10 used on input line 741. (pdftex.def) Requested size: 795.0303pt x 562.1644pt. File: PID.pdf Graphic file (type pdf) - -Package pdftex.def Info: PID.pdf , page11 used on input line 722. + +Package pdftex.def Info: PID.pdf , page10 used on input line 741. (pdftex.def) Requested size: 795.0303pt x 562.1644pt. File: PID.pdf Graphic file (type pdf) - -Package pdftex.def Info: PID.pdf , page11 used on input line 722. + +Package pdftex.def Info: PID.pdf , page10 used on input line 741. (pdftex.def) Requested size: 795.0303pt x 562.1644pt. [53 <./PID.pdf>] pdfTeX warning: /usr/bin/pdflatex (file ./PID.pdf): PDF inclusion: found PDF ve rsion <1.7>, but at most version <1.5> allowed - + File: PID.pdf Graphic file (type pdf) - -Package pdftex.def Info: PID.pdf , page12 used on input line 722. + +Package pdftex.def Info: PID.pdf , page11 used on input line 741. (pdftex.def) Requested size: 795.0303pt x 562.1644pt. File: PID.pdf Graphic file (type pdf) - -Package pdftex.def Info: PID.pdf , page12 used on input line 722. + +Package pdftex.def Info: PID.pdf , page11 used on input line 741. (pdftex.def) Requested size: 795.0303pt x 562.1644pt. File: PID.pdf Graphic file (type pdf) - -Package pdftex.def Info: PID.pdf , page12 used on input line 722. + +Package pdftex.def Info: PID.pdf , page11 used on input line 741. (pdftex.def) Requested size: 795.0303pt x 562.1644pt. [54 <./PID.pdf>] -Package atveryend Info: Empty hook `BeforeClearDocument' on input line 726. - [55] -Package atveryend Info: Empty hook `AfterLastShipout' on input line 726. + +pdfTeX warning: /usr/bin/pdflatex (file ./PID.pdf): PDF inclusion: found PDF ve +rsion <1.7>, but at most version <1.5> allowed + +File: PID.pdf Graphic file (type pdf) + +Package pdftex.def Info: PID.pdf , page12 used on input line 741. +(pdftex.def) Requested size: 795.0303pt x 562.1644pt. +File: PID.pdf Graphic file (type pdf) + +Package pdftex.def Info: PID.pdf , page12 used on input line 741. +(pdftex.def) Requested size: 795.0303pt x 562.1644pt. +File: PID.pdf Graphic file (type pdf) + +Package pdftex.def Info: PID.pdf , page12 used on input line 741. +(pdftex.def) Requested size: 795.0303pt x 562.1644pt. +[55 <./PID.pdf>] +Package atveryend Info: Empty hook `BeforeClearDocument' on input line 745. + [56] +Package atveryend Info: Empty hook `AfterLastShipout' on input line 745. (./document.aux) -Package atveryend Info: Executing hook `AtVeryEndDocument' on input line 726. -Package atveryend Info: Executing hook `AtEndAfterFileList' on input line 726. +Package atveryend Info: Executing hook `AtVeryEndDocument' on input line 745. +Package atveryend Info: Executing hook `AtEndAfterFileList' on input line 745. Package rerunfilecheck Info: File `document.out' has not changed. -(rerunfilecheck) Checksum: 2D9E87FEC6D93F9F3501E4506AD8AD84;9738. +(rerunfilecheck) Checksum: EE0B06B4CF3C74748D09C1D69BD8A1A8;9738. Package logreq Info: Writing requests to 'document.run.xml'. \openout1 = `document.run.xml'. ) Here is how much of TeX's memory you used: - 22537 strings out of 492982 - 356980 string characters out of 6134895 - 847942 words of memory out of 5000000 - 25613 multiletter control sequences out of 15000+600000 + 22553 strings out of 492982 + 357315 string characters out of 6134895 + 846942 words of memory out of 5000000 + 25624 multiletter control sequences out of 15000+600000 12017 words of font info for 47 fonts, out of 8000000 for 9000 1141 hyphenation exceptions out of 8191 45i,18n,67p,2008b,1302s stack positions out of 5000i,500n,10000p,200000b,80000s @@ -1421,10 +1428,10 @@ amsfonts/cm/cmr7.pfb> -Output written on document.pdf (55 pages, 1317438 bytes). +Output written on document.pdf (56 pages, 1411693 bytes). PDF statistics: - 900 PDF objects out of 1000 (max. 8388607) - 786 compressed objects within 8 object streams - 159 named destinations out of 1000 (max. 500000) - 635 words of extra memory for PDF output out of 10000 (max. 10000000) + 913 PDF objects out of 1000 (max. 8388607) + 796 compressed objects within 8 object streams + 163 named destinations out of 1000 (max. 500000) + 640 words of extra memory for PDF output out of 10000 (max. 10000000) diff --git a/document.out b/document.out index 1a2ea60..6ffa8ec 100644 --- a/document.out +++ b/document.out @@ -31,33 +31,33 @@ \BOOKMARK [2][-]{section*.32}{\376\377\000L\000a\000n\000g\000u\000a\000g\000e\000\040\000D\000e\000t\000e\000c\000t\000i\000o\000n}{section*.28}% 31 \BOOKMARK [2][-]{section*.33}{\376\377\000S\000e\000n\000t\000i\000m\000e\000n\000t\000\040\000A\000n\000a\000l\000y\000s\000i\000s}{section*.28}% 32 \BOOKMARK [2][-]{section*.34}{\376\377\000N\000e\000u\000r\000a\000l\000\040\000N\000e\000t\000w\000o\000r\000k}{section*.28}% 33 -\BOOKMARK [2][-]{section*.35}{\376\377\000P\000r\000i\000c\000e\000\040\000F\000o\000r\000e\000c\000a\000s\000t\000i\000n\000g}{section*.28}% 34 -\BOOKMARK [2][-]{section*.36}{\376\377\000F\000r\000o\000n\000t\000e\000n\000d\000\040\000A\000p\000p\000l\000i\000c\000a\000t\000i\000o\000n}{section*.28}% 35 -\BOOKMARK [2][-]{section*.37}{\376\377\000W\000i\000t\000h\000\040\000r\000e\000f\000e\000r\000e\000n\000c\000e\000\040\000t\000o\000\040\000I\000n\000i\000t\000i\000a\000l\000\040\000P\000I\000D}{section*.28}% 36 -\BOOKMARK [2][-]{section*.38}{\376\377\000S\000o\000l\000u\000t\000i\000o\000n\000\040\000S\000u\000m\000m\000a\000r\000y}{section*.28}% 37 -\BOOKMARK [2][-]{section*.39}{\376\377\000D\000a\000t\000a\000\040\000f\000l\000o\000w\000\040\000O\000v\000e\000r\000v\000i\000e\000w}{section*.28}% 38 -\BOOKMARK [1][-]{section*.40}{\376\377\000S\000y\000s\000t\000e\000m\000\040\000D\000e\000s\000i\000g\000n}{}% 39 -\BOOKMARK [2][-]{section*.41}{\376\377\000D\000a\000t\000a\000f\000l\000o\000w\000\040\000D\000e\000s\000i\000g\000n\000s}{section*.40}% 40 -\BOOKMARK [2][-]{section*.42}{\376\377\000U\000M\000L\000\040\000C\000o\000m\000p\000o\000n\000e\000n\000t\000\040\000D\000e\000s\000i\000g\000n}{section*.40}% 41 -\BOOKMARK [2][-]{section*.43}{\376\377\000I\000n\000t\000e\000r\000f\000a\000c\000e\000\040\000D\000e\000s\000i\000g\000n}{section*.40}% 42 -\BOOKMARK [1][-]{section*.44}{\376\377\000I\000m\000p\000l\000e\000m\000e\000n\000t\000a\000t\000i\000o\000n}{}% 43 -\BOOKMARK [2][-]{section*.45}{\376\377\000D\000a\000t\000a\000\040\000c\000o\000l\000l\000e\000c\000t\000i\000o\000n}{section*.44}% 44 -\BOOKMARK [3][-]{section*.46}{\376\377\000P\000r\000i\000c\000e\000\040\000T\000i\000m\000e\000-\000s\000e\000r\000i\000e\000s\000\040\000D\000a\000t\000a}{section*.45}% 45 -\BOOKMARK [2][-]{section*.47}{\376\377\000D\000a\000t\000a\000\040\000p\000r\000o\000c\000e\000s\000s\000i\000n\000g}{section*.44}% 46 -\BOOKMARK [3][-]{section*.48}{\376\377\000P\000r\000e\000p\000r\000o\000c\000e\000s\000s\000i\000n\000g}{section*.47}% 47 -\BOOKMARK [3][-]{section*.52}{\376\377\000S\000p\000a\000m\000\040\000F\000i\000l\000t\000e\000r\000i\000n\000g}{section*.47}% 48 -\BOOKMARK [2][-]{section*.55}{\376\377\000S\000e\000n\000t\000i\000m\000e\000n\000t\000\040\000A\000n\000a\000l\000y\000s\000i\000s}{section*.44}% 49 -\BOOKMARK [3][-]{section*.56}{\376\377\000V\000A\000D\000E\000R}{section*.55}% 50 -\BOOKMARK [2][-]{section*.57}{\376\377\000R\000e\000c\000u\000r\000r\000e\000n\000t\000\040\000N\000e\000u\000r\000a\000l\000\040\000N\000e\000t\000w\000o\000r\000k\000\040\000-\000\040\000L\000S\000T\000M}{section*.44}% 51 -\BOOKMARK [3][-]{section*.58}{\376\377\000T\000r\000a\000i\000n\000i\000n\000g\000\040\000a\000n\000d\000\040\000T\000e\000s\000t\000i\000n\000g\000\040\000M\000o\000d\000e\000l}{section*.57}% 52 -\BOOKMARK [3][-]{section*.59}{\376\377\000S\000c\000o\000r\000i\000n\000g\000\040\000a\000n\000d\000\040\000V\000a\000l\000i\000d\000a\000t\000i\000o\000n}{section*.57}% 53 -\BOOKMARK [3][-]{section*.60}{\376\377\000F\000u\000t\000u\000r\000e\000\040\000P\000r\000e\000d\000i\000c\000t\000i\000o\000n\000\040\000F\000o\000r\000e\000c\000a\000s\000t\000i\000n\000g}{section*.57}% 54 -\BOOKMARK [1][-]{section*.61}{\376\377\000T\000e\000s\000t\000i\000n\000g\000:\000\040\000V\000e\000r\000i\000f\000i\000c\000a\000t\000i\000o\000n\000\040\000a\000n\000d\000\040\000R\000e\000f\000l\000e\000c\000t\000i\000o\000n}{}% 55 -\BOOKMARK [1][-]{section*.62}{\376\377\000D\000i\000s\000c\000u\000s\000s\000i\000o\000n\000:\000\040\000C\000o\000n\000t\000r\000i\000b\000u\000t\000i\000o\000n\000\040\000a\000n\000d\000\040\000R\000e\000f\000l\000e\000c\000t\000i\000o\000n}{}% 56 -\BOOKMARK [2][-]{section*.63}{\376\377\000L\000i\000m\000i\000t\000a\000t\000i\000o\000n\000s}{section*.62}% 57 -\BOOKMARK [1][-]{section*.64}{\376\377\000C\000o\000n\000c\000l\000u\000s\000i\000o\000n\000\040\000a\000n\000d\000\040\000F\000u\000t\000u\000r\000e\000\040\000I\000m\000p\000r\000o\000v\000e\000m\000e\000n\000t\000s}{}% 58 -\BOOKMARK [2][-]{section*.65}{\376\377\000C\000o\000n\000c\000l\000u\000s\000i\000o\000n}{section*.64}% 59 -\BOOKMARK [2][-]{section*.66}{\376\377\000F\000u\000t\000u\000r\000e\000\040\000I\000m\000p\000r\000o\000v\000e\000m\000e\000n\000t\000s}{section*.64}% 60 -\BOOKMARK [1][-]{section*.68}{\376\377\000A\000p\000p\000e\000n\000d\000i\000c\000e\000s}{}% 61 -\BOOKMARK [2][-]{section*.69}{\376\377\000A\000p\000p\000e\000n\000d\000i\000x\000\040\000A\000\040\000-\000\040\000P\000r\000o\000j\000e\000c\000t\000\040\000I\000n\000i\000t\000i\000a\000t\000i\000o\000n\000\040\000D\000o\000c\000u\000m\000e\000n\000t}{section*.68}% 62 -\BOOKMARK [2][-]{section*.70}{\376\377\000A\000p\000p\000e\000n\000d\000i\000x\000\040\000B\000\040\000-\000\040\000L\000o\000g\000\040\000b\000o\000o\000k}{section*.68}% 63 +\BOOKMARK [2][-]{section*.36}{\376\377\000P\000r\000i\000c\000e\000\040\000F\000o\000r\000e\000c\000a\000s\000t\000i\000n\000g}{section*.28}% 34 +\BOOKMARK [2][-]{section*.37}{\376\377\000F\000r\000o\000n\000t\000e\000n\000d\000\040\000A\000p\000p\000l\000i\000c\000a\000t\000i\000o\000n}{section*.28}% 35 +\BOOKMARK [2][-]{section*.38}{\376\377\000W\000i\000t\000h\000\040\000r\000e\000f\000e\000r\000e\000n\000c\000e\000\040\000t\000o\000\040\000I\000n\000i\000t\000i\000a\000l\000\040\000P\000I\000D}{section*.28}% 36 +\BOOKMARK [2][-]{section*.39}{\376\377\000S\000o\000l\000u\000t\000i\000o\000n\000\040\000S\000u\000m\000m\000a\000r\000y}{section*.28}% 37 +\BOOKMARK [2][-]{section*.40}{\376\377\000D\000a\000t\000a\000\040\000f\000l\000o\000w\000\040\000O\000v\000e\000r\000v\000i\000e\000w}{section*.28}% 38 +\BOOKMARK [1][-]{section*.41}{\376\377\000S\000y\000s\000t\000e\000m\000\040\000D\000e\000s\000i\000g\000n}{}% 39 +\BOOKMARK [2][-]{section*.42}{\376\377\000D\000a\000t\000a\000f\000l\000o\000w\000\040\000D\000e\000s\000i\000g\000n\000s}{section*.41}% 40 +\BOOKMARK [2][-]{section*.43}{\376\377\000U\000M\000L\000\040\000C\000o\000m\000p\000o\000n\000e\000n\000t\000\040\000D\000e\000s\000i\000g\000n}{section*.41}% 41 +\BOOKMARK [2][-]{section*.44}{\376\377\000I\000n\000t\000e\000r\000f\000a\000c\000e\000\040\000D\000e\000s\000i\000g\000n}{section*.41}% 42 +\BOOKMARK [1][-]{section*.46}{\376\377\000I\000m\000p\000l\000e\000m\000e\000n\000t\000a\000t\000i\000o\000n}{}% 43 +\BOOKMARK [2][-]{section*.47}{\376\377\000D\000a\000t\000a\000\040\000c\000o\000l\000l\000e\000c\000t\000i\000o\000n}{section*.46}% 44 +\BOOKMARK [3][-]{section*.48}{\376\377\000P\000r\000i\000c\000e\000\040\000T\000i\000m\000e\000-\000s\000e\000r\000i\000e\000s\000\040\000D\000a\000t\000a}{section*.47}% 45 +\BOOKMARK [2][-]{section*.49}{\376\377\000D\000a\000t\000a\000\040\000p\000r\000o\000c\000e\000s\000s\000i\000n\000g}{section*.46}% 46 +\BOOKMARK [3][-]{section*.50}{\376\377\000P\000r\000e\000p\000r\000o\000c\000e\000s\000s\000i\000n\000g}{section*.49}% 47 +\BOOKMARK [3][-]{section*.54}{\376\377\000S\000p\000a\000m\000\040\000F\000i\000l\000t\000e\000r\000i\000n\000g}{section*.49}% 48 +\BOOKMARK [2][-]{section*.57}{\376\377\000S\000e\000n\000t\000i\000m\000e\000n\000t\000\040\000A\000n\000a\000l\000y\000s\000i\000s}{section*.46}% 49 +\BOOKMARK [3][-]{section*.58}{\376\377\000V\000A\000D\000E\000R}{section*.57}% 50 +\BOOKMARK [2][-]{section*.59}{\376\377\000R\000e\000c\000u\000r\000r\000e\000n\000t\000\040\000N\000e\000u\000r\000a\000l\000\040\000N\000e\000t\000w\000o\000r\000k\000\040\000-\000\040\000L\000S\000T\000M}{section*.46}% 51 +\BOOKMARK [3][-]{section*.60}{\376\377\000T\000r\000a\000i\000n\000i\000n\000g\000\040\000a\000n\000d\000\040\000T\000e\000s\000t\000i\000n\000g\000\040\000M\000o\000d\000e\000l}{section*.59}% 52 +\BOOKMARK [3][-]{section*.61}{\376\377\000S\000c\000o\000r\000i\000n\000g\000\040\000a\000n\000d\000\040\000V\000a\000l\000i\000d\000a\000t\000i\000o\000n}{section*.59}% 53 +\BOOKMARK [3][-]{section*.62}{\376\377\000F\000u\000t\000u\000r\000e\000\040\000P\000r\000e\000d\000i\000c\000t\000i\000o\000n\000\040\000F\000o\000r\000e\000c\000a\000s\000t\000i\000n\000g}{section*.59}% 54 +\BOOKMARK [1][-]{section*.63}{\376\377\000T\000e\000s\000t\000i\000n\000g\000:\000\040\000V\000e\000r\000i\000f\000i\000c\000a\000t\000i\000o\000n\000\040\000a\000n\000d\000\040\000R\000e\000f\000l\000e\000c\000t\000i\000o\000n}{}% 55 +\BOOKMARK [1][-]{section*.64}{\376\377\000D\000i\000s\000c\000u\000s\000s\000i\000o\000n\000:\000\040\000C\000o\000n\000t\000r\000i\000b\000u\000t\000i\000o\000n\000\040\000a\000n\000d\000\040\000R\000e\000f\000l\000e\000c\000t\000i\000o\000n}{}% 56 +\BOOKMARK [2][-]{section*.65}{\376\377\000L\000i\000m\000i\000t\000a\000t\000i\000o\000n\000s}{section*.64}% 57 +\BOOKMARK [1][-]{section*.66}{\376\377\000C\000o\000n\000c\000l\000u\000s\000i\000o\000n\000\040\000a\000n\000d\000\040\000F\000u\000t\000u\000r\000e\000\040\000I\000m\000p\000r\000o\000v\000e\000m\000e\000n\000t\000s}{}% 58 +\BOOKMARK [2][-]{section*.67}{\376\377\000C\000o\000n\000c\000l\000u\000s\000i\000o\000n}{section*.66}% 59 +\BOOKMARK [2][-]{section*.68}{\376\377\000F\000u\000t\000u\000r\000e\000\040\000I\000m\000p\000r\000o\000v\000e\000m\000e\000n\000t\000s}{section*.66}% 60 +\BOOKMARK [1][-]{section*.70}{\376\377\000A\000p\000p\000e\000n\000d\000i\000c\000e\000s}{}% 61 +\BOOKMARK [2][-]{section*.71}{\376\377\000A\000p\000p\000e\000n\000d\000i\000x\000\040\000A\000\040\000-\000\040\000P\000r\000o\000j\000e\000c\000t\000\040\000I\000n\000i\000t\000i\000a\000t\000i\000o\000n\000\040\000D\000o\000c\000u\000m\000e\000n\000t}{section*.70}% 62 +\BOOKMARK [2][-]{section*.72}{\376\377\000A\000p\000p\000e\000n\000d\000i\000x\000\040\000B\000\040\000-\000\040\000L\000o\000g\000\040\000b\000o\000o\000k}{section*.70}% 63 diff --git a/document.pdf b/document.pdf index 13ed0fc..98fc215 100644 Binary files a/document.pdf and b/document.pdf differ diff --git a/document.synctex.gz b/document.synctex.gz index aee2c6a..c7077dd 100644 Binary files a/document.synctex.gz and b/document.synctex.gz differ diff --git a/document.tex b/document.tex index 84c00aa..8aa739e 100644 --- a/document.tex +++ b/document.tex @@ -48,7 +48,7 @@ \large School of Mathematical, Physical and Computational Sciences - + \newline Indivdual Project - CS3IP16 \end{center}} @@ -122,22 +122,22 @@ \begin{center} \section{Introduction}\label{introduction} \end{center} - The premise of this project is to investigate into whether the sentiment expressed in social media has a correlation to the prices of cryptocurrencies and how this could be used to predict future changes in the price. + The premise of this project is to investigate whether the sentiment expressed in social media has a correlation to the prices of cryptocurrencies and how this could be used to predict future changes in the price. - The chosen cryptocurrency that will be of focus for this project will be the currency Bitcoin (BTC), due to having the largest community and backing and has been known to lead other fiat currencies. Bitcoin is seen as one, if not the first cryptocurrency to bring a wider following to the peer-to-peer token transaction scene since 2009. Although it was not the first token to utilise blockchain technology, it allowed investors to openly trade a public cryptocurrency which provided pseudonymous means of transferring funds through the internet. Thus it has been around longer than most of the other fiat currencies and is the most popular crypto-token due to it's larger community base. + The chosen cryptocurrency that will be of focus for this project will be the currency Bitcoin (BTC), due to having the largest community and backing and has been known to lead other fiat currencies. Bitcoin is seen as one, if not the first cryptocurrency to bring a broader following to the peer-to-peer token transaction scene since 2009. Although it was not the first token to utilise blockchain technology, it allowed investors to openly trade a public cryptocurrency which provided pseudonymous means of transferring funds through the internet. Thus it has been around longer than most of the other fiat currencies and is the most popular crypto-token due to it's more extensive community base. - Most financial commodities are subject to the whim of public confidence and are the core of it's base value. A platform that is frequently used for the public to convey their opinions on a commodity is that of Twitter which provides arguably biased information and opinions. Whether the opinions present a basis in facts or not, they are usually taken at face value and can influence the public opinion of given topics. As Bitcoin has been around since 2009 the opinions and information on the commodity are prevalent through the platform. - In the paper \textit{Sentiment Analysis of Twitter Data for Predicting Stock Market Movements} by \textit{Majhi et al.} \cite{1} 2.5 million tweets on Microsoft were extracted from Twitter, sentiment analysis and logistical regression performed on the data yielded 69.01\% accuracy for a 3-day period on the increase/decrease in stock price. These results showed a "\textit{good correlation between stock market movements and the sentiments of public expressed in Twitter}". + Most financial commodities are subject to the whim of public confidence and are the core of its base value. A platform that is frequently used for the public to convey their opinions on a commodity is that of Twitter which provides arguably biased information and opinions. Whether the opinions present a basis in facts or not, they are usually taken at face value and can influence the public opinion of given topics. As Bitcoin has been around since 2009 the opinions and information on the commodity are prevalent through the platform. + In the paper \textit{Sentiment Analysis of Twitter Data for Predicting Stock Market Movements} by \textit{Majhi et al.} \cite{1} 2.5 million tweets on Microsoft were extracted from Twitter, sentiment analysis and logistical regression performed on the data yielded 69.01\% accuracy for a 3-day period on the increase/decrease in stock price. These results showed a "\textit{good correlation between stock market movements and the sentiments of the public expressed in Twitter}". The background of this project is in response to the volatility of the cryptocurrency market, which can fluctuate at a moments notice and can be seen to be social media driven. The history of the price of Bitcoin and what was being discussed on the currency around it's most volatile period to-date, Nov-2017 to Feb-2018, shows a strong bullish trend which saw Bitcoin reach a \$19,500 high in mid-Dec. While social media, such as Twitter, during that period was had an extremely positive outlook on the cryptocurrency. The trend was short lived and saw the market crash only a month later, with only a couple of sell-offs, expected for the holidays rush, accompanied by negative outlooks posted on social media turned the market against itself which saw the longest bearish market in Bitcoin's history and is still trying to recover today. Due to how volatile the crypto-market can be, there is a need to either mitigate or to anticipate where the markets are heading. As the crypto-market and Bitcoin are affected by socially constructed opinions, either through Twitter, news articles or other forms of media, there is a way to perform the latter, where the prices of Bitcoin could be predicted based on the sentiment gathered from social media outlets. - The aim of this project is to create a tool that gathers tweets from Twitter, obtains the overall sentiment score of the given text while gathering historical price data for the time period gathering occurs. Features are then extracted from the gathered data and used in a neural network to ascertain whether the price of the currency can be predicted from the correlation between the sentiment and price history of the data. + This project aims to create a tool that gathers tweets from Twitter, obtains the overall sentiment score of the given text while gathering historical price data for the period gathering occurs. Features are then extracted from the gathered data and used in a neural network to ascertain whether the price of the currency can be predicted from the correlation between the sentiment and price history of the data. This report will discuss the justifications for the project and the problems it will be attempting to resolve, the stakeholders that would benefit the most from this system and what this project will not attempt to accomplish. Similar tools will be critiqued and examined for their feature set and credibility in the literature review along with current sentiment analysers, algorithms, natural language processing techniques and neural networks in their respective topics and comparing their accuracy for this project purpose. The solution approach will discuss the decisions and reasoning behind choosing the techniques and tools used for this project and will outline the requirements for this project. - Implementation of the chosen techniques and tools, with the discussion of important functions of the system will formulate the implementation section of this report with an in-detail explanation of the function's use and data flow of the system. + Implementation of the chosen techniques and tools, with the discussion of essential functions of the system will formulate the implementation section of this report with an in-detail explanation of the function's use and data flow of the system. \newpage @@ -147,11 +147,11 @@ \subsection{Problem Statement}\label{statement} - The key problems this project attempts to address are that of, an open-source system available to the public that aids in the analysis and prediction of BTC. The accuracy of open-source tools and technology when applied to the trading market scene and to identify whether there is a correlation between Twitter sentiment and BTC price fluctuation. While there are existing tools only a few are available to the public and only provide basic functionality, while others are kept in-house of major corporations who invest into this problem domain. + The fundamental problems this project attempts to address are that of, an open-source system available to the public that aids in the analysis and prediction of BTC. The accuracy of open-source tools and technology when applied to the trading market scene and to identify whether there is a correlation between Twitter sentiment and BTC price fluctuation. While there are existing tools, only a few are available to the public and only provide basic functionality, while others are kept in-house of major corporations who invest in this problem domain. - The other issue presented here is that assuming perfect accuracy can be achieved is naive. As this project will only be using existing tools and technologies thus, there are limitations to the accuracy of what can be obtained. One of that being the suitability of the tools, there are no open-source sentiment analysers for stock market prediction, thus finding a specifically trained analyser for the chosen domain in highly unlikely. In relation, finding the most suitable machine learning or neural network is equally important as this will determine the accuracy of the predictions. Due to being a regression problem, machine learning techniques and neural networks that focus around this and forecasting should be considered. + The other issue presented here is that assuming perfect accuracy can be achieved is naive. As this project will only be using existing tools and technologies; thus, there are limitations to the accuracy of what can be obtained. One of that being the suitability of the tools, there are no open-source sentiment analysers for stock market prediction, thus finding a specifically trained analyser for the chosen domain in highly unlikely. In relation, finding the most suitable machine learning or neural network is equally important as this will determine the accuracy of the predictions. Due to being a regression problem, machine learning techniques and neural networks that focus around this and forecasting should be considered. - The accuracy and suitability of various machine learning methods and neural networks are a known issue in their respective domains, this investigation should be carried out to determine their suitability for their needed use in this project and will be detailed in the literature review. + The accuracy and suitability of various machine learning methods and neural networks are a known issue in their respective domains. This investigation should be carried out to determine their suitability for their needed use in this project and will be detailed in the literature review. This project will focus on the investigation of these technologies and tools to justify whether it is feasible to predict the price of BTC based on historical price and the sentiment gathered from Twitter. Limitations of the system and it's accuracy in predictions should be investigated and discussed to determine the implemented solution is the more suitable compared to other methods. @@ -171,7 +171,7 @@ \subsection{Project Motivation} The motivation behind the project stems from a range of points, from personal and public issues with the volatility if the crypto-market, and how losses specifically could be mitigated. The personal motivations behind the conceptualisation of this began two years ago during the crash of late 2017-2018, which saw new investors blindly jump into the trend that was buying cryptocurrencies. During this period of November to December 2017 saw Bitcoin's price reach \$20,000 from \$5,000, new public investors jumped on the chance to buy into the trend of possibly making quick profits and the fear of missing out (FOMO). In late December, a few holiday sell-offs occurred from business and big investors, this coupled with a few negative outlooks posted on social media by news outlets caused the market to implode causing investors to panic sell one after the other and posting negativity on social, thus causing more decline in the market. As a result, this caused personal monetary loss and distress as being a long-term investor. - Another motivation is that at the time of writing, there are no pubically available systems that combine sentiment analysis with historical price to forecast the price of Bitcoin or any other cryptocurrency. There are papers and a few code repositories that implement a similar concepts \cite{2} - \textit{Use of a Multi-layer Perceptron network for moving averages in Bitcoin price}, \cite{3} - \textit{Predicting Bitcoin price fluctuation with Twitter sentiment analysis}, \cite{4} - \textit{Predict Tomorrows Bitcoin (BTC) Price with Recurrent Neural Networks} but are not operational. System such as \cite{1} hosted on Coingecko, a popular cryptocurrency track site, provides a tool for basic sentiment analysis but doesn't give an evaluated indication of the direction of the market as a prediction. This leaves the public to the whim of volatility of the market without a means to know what the next, say an hour, could entail to possibly reduce losses if the market drops. Such system are usually kept in-house of major corporations whom invest significant time into tackling such a problem. Additionly, this could be seen as a positive for major investors, as such a system could cause panic selling if public investors soley trusted such a system. + Another motivation is that at the time of writing, there are no publically available systems that combine sentiment analysis with a historical price to forecast the price of Bitcoin or any other cryptocurrency. There are papers and a few code repositories that implement a similar concepts \cite{2} - \textit{Use of a Multi-layer Perceptron network for moving averages in Bitcoin price}, \cite{3} - \textit{Predicting Bitcoin price fluctuation with Twitter sentiment analysis}, \cite{4} - \textit{Predict Tomorrows Bitcoin (BTC) Price with Recurrent Neural Networks} but are not operational. A system such as \cite{1} hosted on Coingecko, a popular cryptocurrency track site, provides a tool for basic sentiment analysis but doesn't give an evaluated indication of the direction of the market as a prediction. This leaves the public to the whim of volatility of the market without a means to know what the next, say an hour, could entail to possibly reduce losses if the market drops. Such systems are usually kept in-house of major corporations who invest significant time into tackling such a problem. Additionlly, this could be seen as a positive for major investors, as such a system could cause panic selling if public investors solely trusted such a system. \newpage \subsection{Technical Specification} @@ -248,22 +248,22 @@ \subsection{Related research} - There has been a plentiful amount of research conducted in this problem domain. Numerous theses globally have been published in recent years on the topic of cryptocurrency market predictions and analysis, and even more, research conducted on general stock markets further back. + There has been an abundant amount of research conducted in this problem domain. Many theses globally have been published in recent years on the topic of cryptocurrency market predictions and analysis, and even more, research conducted on general stock markets further back. - The thesis written by \textit{Evita Stenqvist and Jacob Lonno} of the \textit{KTH Royal Institute of Technology} \cite{3} investigates the use of sentiment expressed through micro-blogging such as Twitter can have on the price fluctuations of Bitcoin. Its primary focus was creating an analyser for the sentiment of tweets more accurately \textit{"by not only taking into account negation, but also valence, common slang and smileys"}, than that of former researchers that \textit{"mused that accounting for negations in text may be a step in the direction of more accurate predictions."}. This would be built upon the lexicon-based sentiment analyser VADER to ascertain the overall sentiment scores were grouped into time-series for each interval from 5 minutes to 4 hours, along with the interval prices for Bitcoin. The model chosen was a naive binary classified vectors of predictions for a certain threshold to \textit{"ultimately compare the predictions to actual historical price data"}. The results of this research suggest that a binary classification model of varying threshold over time-shifts in time-series data was "lackluster", seeing the number of predictions decreasing rapidly as the threshold changed. This research is a good basis of starting research upon, as it suggests tools such as VADER for sentiment analysis and that the use of a machine learning algorithm would be a next step in the project that would yield better more accurate results. + The thesis written by \textit{Evita Stenqvist and Jacob Lonno} of the \textit{KTH Royal Institute of Technology} \cite{3} investigates the use of sentiment expressed through micro-blogging such as Twitter can have on the price fluctuations of Bitcoin. Its primary focus was creating an analyser for the sentiment of tweets more accurately \textit{"by not only taking into account negation, but also valence, common slang and smileys"}, than that of former researchers that \textit{"mused that accounting for negations in text may be a step in the direction of more accurate predictions."}. This would be built upon the lexicon-based sentiment analyser VADER to ascertain the overall sentiment scores were grouped into time-series for each interval from 5 minutes to 4 hours, along with the interval prices for Bitcoin. The model chosen was a naive binary classified vectors of predictions for a certain threshold to \textit{"ultimately compare the predictions to actual historical price data"}. The results of this research suggest that a binary classification model of varying threshold over time-shifts in time-series data was "lackluster", seeing the number of predictions decreasing rapidly as the threshold changed. This research is a reasonable basis of starting research upon, as it suggests tools such as VADER for sentiment analysis and that the use of a machine learning algorithm would be a next step in the project that would yield better more accurate results. Another thesis written by \textit{Pagolu, Venkata Sasank and Reddy Kamal Nayan, Panda Ganapati and Majhi, Babita} \cite{1} on "Sentiment Analysis of Twitter Data for Predicting Stock Market Movements" 2.5 million tweets on Microsoft were extracted from Twitter, sentiment analysis and logistical regression performed on the data yielded 69.01\% accuracy for a 3-day period on the increase/decrease in stock price. These results showed a "\textit{good correlation between stock market movements and the sentiments of the public expressed in Twitter}". Using various natural language pre-processing tweets for feature extraction such as N-gram representation the sentiment from tweets were extrapolated. Both Word2vec and a random forest classifier were compared for accuracy, Word2vec being chosen over the machine learning model. Word2vec, being a group of related shallow two-layer neural network models to produce word embeddings. - A topic that reoccurs in various papers and theses is that of the use and focus of regression techniques and machine learning methods. Few implement a fully fledged neural network, the above paper attempts to use a simple network to achieve predictions of classification of sentiment for stock market movement then correlated this with historical data of prices. An article posted on "Code Project" by Intel Corporation \cite{5} compares the accuracy of three machine learning algorithms; Random Forest, Logistic Regression and Multi-Layer Perceptron (MLP) classifiers on predicting the price fluctuations of Bitcoin with embedded price indices. Results showing \textit{"that using the MLP classifier (a.k.a. neural networks) showed better results than logistic regression and random forest trained models"}. This assumption can be backed up by the results from a thesis posted on IEEE \cite{6} which compares a Bayesian optimised recurrent neural network and a Long Short Term Memory (LSTM) network. Showing the LSTM model achieving \textit{"the highest classification accuracy of 52\% and a RMSE of 8\%"}. With an interest in neural networks personally and with little papers utilising them for this purpose a neural network will thus be implemented, and the accuracy of one's predictions with use of sentiment analysis data analysed and discussed. + A topic that reoccurs in various papers and theses is that of the use and focus of regression techniques and machine learning methods. Few implement a fully fledged neural network; the above paper attempts to use a simple network to achieve predictions of classification of sentiment for stock market movement then correlated this with historical data of prices. An article posted on "Code Project" by Intel Corporation \cite{5} compares the accuracy of three machine learning algorithms; Random Forest, Logistic Regression and Multi-Layer Perceptron (MLP) classifiers on predicting the price fluctuations of Bitcoin with embedded price indices. Results showing \textit{"that using the MLP classifier (a.k.a. neural networks) showed better results than logistic regression and random forest trained models"}. This assumption can be backed up by the results from a thesis posted on IEEE \cite{6} which compares a Bayesian optimised recurrent neural network and a Long Short Term Memory (LSTM) network - showing the LSTM model achieving \textit{"the highest classification accuracy of 52\% and a RMSE of 8\%"}. With interest in neural networks personally and with little papers utilising them for this purpose a neural network will thus be implemented, and the accuracy of one's predictions with use of sentiment analysis data analysed and discussed. \subsection{Data Collection}\label{tweet_collection} \subsubsection{Twitter and Twitter API} - Twitter is a micro-blogging platform that was launched in 2006 and provides it's users the ability to publish short messages of 140 characters. The messages published could be of any form, from news snippets, advertisement, or the prevalent publication of opinions which allowed a platform of extensive diversity and knowledge wealth. As of the time of writing, the message character limit was increased to 280 characters, the platform has over 300 million monthly active users and around 1 million tweets are published per day. Due to the length restriction and the primary use of the platform to express opinions Twitter is seen as a gold mine for opinion mining. + Twitter is a micro-blogging platform that was launched in 2006 and provides it's users the ability to publish short messages of 140 characters. The messages published could be of any form, from news snippets, advertisement, or the prevalent publication of opinions which allowed a platform of extensive diversity and knowledge wealth. As of the time of writing, the message character limit was increased to 280 characters, the platform has over 300 million monthly active users, and around 1 million tweets are published per day. Due to the length restriction and the primary use of the platform to express opinions Twitter is seen as a gold mine for opinion mining. - The Twitter API has an extensive range of endpoints that provide access from streaming tweets for a given hashtag, obtaining historical tweets for a given time-period and hashtag, posting tweets on a users account and to change settings on a user account with authentication. The exhaustive range of features provided by these endpoints makes data collection from Twitter straight forward as one can target a specific endpoint for the required data. Due to Twitter being the target for opinion mining within this project the Twitter API will ultimately need to be utilised. This can either be used for the gathering of historical tweets or streaming current tweets for the \#Bitcoin hashtag. + The Twitter API has an extensive range of endpoints that provide access from streaming tweets for a given hashtag, obtaining historical tweets for a given time-period and hashtag, posting tweets on a users account and to change settings on a user account with authentication. The exhaustive range of features provided by these endpoints makes data collection from Twitter straight forward as one can target a specific endpoint for the required data. Due to Twitter being the target for opinion mining within this project the Twitter API will ultimately need to be utilised. This can either be used for the gathering of historical tweets or streaming current tweets for the \#Bitcoin hashtag. - There are, however, limitations and rate limits imposed on users of the API. Twitter employs a tiering system for the API - Standard, Premium and Enterprise tiers, each of which provides different amounts of access for data collection. If the API were used to capture historical data for a span of 3 months, each tier is allowed to obtain varying amounts of data for different durations; \cite{7} + There are, however, limitations and rate limits imposed on users of the API. Twitter employs a tiering system for the API - Standard, Premium and Enterprise tiers, each of which provides different amounts of access for data collection. If the API were to be used to capture historical data for a span of 3 months, each tier is allowed to obtain varying amounts of data for different durations; \cite{7} \begin{itemize} \item A Standard user would be able to capture 100 recent tweets for the past 7 days @@ -281,7 +281,7 @@ In short, sentiment analysis is the process and discovery of computationally identifying and categorising the underlining opinions and subjectivity expressed in written language. This process determines the writer's attitude towards a particular topic as either being positive, neutral or negative in terms of opinion, known as polarity classification. \subsubsection{Natural Language Processing}\label{algorithms} - Polarity classification is the focus of sentiment analysis and is a well-known problem in natural language processing that has had significant attention by researchers in recent years \cite{1}\cite{3}\cite{6}\cite{10}. Traditional approaches to this have usually been classified to dictionary-based approaches that use a pre-constructed sentiment lexicons such as VADER or usually confined to machine learning approaches. The later requires an extensive amount of natural language pre-processing to extrapolate vectors and features from given text, this is then fed into a machine learning classifier which attempts to categorise words to a level of sentiment polarity. Natural language pre-processing techniques, supported by the NLTK (Natural Language Toolkit) python package , that would be required for this approach would consist of; + Polarity classification is the focus of sentiment analysis and is a well-known problem in natural language processing that has had significant attention by researchers in recent years \cite{1}\cite{3}\cite{6}\cite{10}. Traditional approaches to this have usually been classified to dictionary-based approaches that use pre-constructed sentiment lexicons such as VADER or usually confined to machine learning approaches. The latter requires an extensive amount of natural language pre-processing to extrapolate vectors and features from the given text; this is then fed into a machine learning classifier which attempts to categorise words to a level of sentiment polarity. Natural language pre-processing techniques, supported by the NLTK (Natural Language Toolkit) python package that would be required for this approach would consist of; \begin{itemize} \item Tokenisation: The act of splitting a stream of text into smaller units of typographical tokens which isolate unneeded punctuation. @@ -292,17 +292,17 @@ \item Ngrams: ... \end{itemize} - The former, seen and has been proven to provide higher accuracy than traditional machine learning approaches \cite{11}, and need little pre-proeccesing conducted on the data as words have a pre-defined sentiment classification in a provided lexicon. Although these lexicons can be complex to create, they generally require little resources to use and add to. + The former, seen and has been proven to provide higher accuracy than traditional machine learning approaches \cite{11}, and need little pre-processing conducted on the data as words have a pre-defined sentiment classification in a provided lexicon. Although these lexicons can be complex to create, they generally require little resources to use and alter. \subsubsection{Valence Aware Dictionary and sEntiment Reasoning}\label{Vader} - VADER is a combined lexicon and rule-based sentiment analysis tool that is specifically attuned to sentiments expressed in social media, and works well on texts from other domains. It is capable of detecting the polarity of a given text - positivity, neurality, and negativity \cite{12}. VADER uses a human-centric approach to sentiment analysis, combining qualitative analysis and empirical validation by using human raters to rate level of sentiment for words in it’s lexicon. Vader also has emoticon support which maps these colloquailisms have pre-defined intensities in its lexicon, which makes VADER specifically suitable for the social media domain where the used of emoticons, utf-8 emojis and slang such as "Lol" and "Yolo" are prevalent within text. Additionally, VADER is provided as a lexicon and a python library under the MIT license, thus means that it is open-source software. This means that the lexicon can be altered and added to making it able to being tailored to specific topic domains. + VADER is a combined lexicon and rule-based sentiment analysis tool that is specifically attuned to sentiments expressed in social media and works well on texts from other domains. It is capable of detecting the polarity of a given text - positivity, neutrality, and negativity \cite{12}. VADER uses a human-centric approach to sentiment analysis, combining qualitative analysis and empirical validation by using human raters to rate the level of sentiment for words in its lexicon. Vader also has emoticon support which maps these colloquialisms have pre-defined intensities in its lexicon, which makes VADER specifically suitable for the social media domain were the use of emoticons, utf-8 emojis and slang such as "Lol" and "Yolo" are prevalent within the text. Additionally, VADER is provided as a lexicon and a python library under the MIT license, this means that it is open-source software. This means that the lexicon can be altered and added to abling it to be tailored to specific topic domains. VADER was constructed by examining and extracting features from three pre-existing well-established and human-validated sentiment lexicons \cite{12} - (LIWC) Linguistic Inquiry and Word Count, (ANEW) Affective Norms for English Words, and (GI) General Inquirer. This is supplemented with additional lexicon features \textit{"commonly used to express sentiment in social media text (emoticons, acronyms and slang)"} \cite{12} and uses "wisdom-of-the-crowd" approach \cite{13} to establish a point of estimations of sentiment valance for each lexical feature candidate. This was evaluated for the impact of grammatical and syntactical rules and 7,500+ lexical features, with mean valence \textit{"<> zero, and SD <= 2.5"} as a human-validated "gold-standard" sentiment lexicon. \cite{12}\textit{Section 3.1} VADER is seen as a "Gold Standard" for sentiment analysis, in the paper for VADER, \cite{12} \textit{A Parsimonious Rule-based Model for Sentiment Analysis of Social Media Text}, it was compared against 11 other \textit{"highly regarded sentiment analysis tools/techniques on a corpus of over 4.2K tweets"} for polarity classification across 4 domains. Results showing VADER, across Social media text, Amazon reviews, movie reviews and Newspaper editorials, consistently outperforming other sentiment analysis tools and techniques showing a particular trend in performing significantly higher on analysis of sentiment in tweets. \cite{12} \textit{Section 4: Results} \subsection{Neural Networks}\label{networks} - A neural network is a set of perceptrons modelled loosely after the human brain that is designed to recognise patterns in whatever domain it is intended to be trained on. A neural network can consist of multiple machine perceptrons or clustering layers in a large mesh network and the patterns they recognise are numerical which are contained in vectors. Pre-processed data, confined and processed into pre-defined vector labels, are used to teach a neural network for a given task. While this differs from how an algorithm is coded to a particular task, neural networks cannot be programmed directly for the task. The requirement is for them to learn from the information by use of different learning strategies; \cite{14}\cite{15} + A neural network is a set of perceptrons modelled loosely after the human brain that is designed to recognise patterns in whatever domain it is intended to be trained using. A neural network can consist of multiple machine perceptrons or clustering layers in a large mesh network, and the patterns they recognise are numerical which are contained in vectors. Pre-processed data, confined and processed into pre-defined vector labels, are used to teach a neural network for a given task. While this differs from how an algorithm is coded to a particular task, neural networks cannot be programmed directly for the task. The requirement is for them to learn from the information by use of different learning strategies; \cite{14}\cite{15} \begin{center} \includegraphics[width=10cm,height=6cm]{images/perceptron.png} @@ -317,12 +317,12 @@ \end{itemize} \subsubsection{Recurrent Neural Network (RNN)}\label{types} - The type of neural network that is of focus for this project will be that of a Long-Short Term Memory (LSTM), however, it is important to understand how this is an extension of a Recurrent Neural Network (RNN) and how the underlying network works. + The type of neural network that is of focus for this project will be that of a Long-Short Term Memory (LSTM); however, it is crucial to understand how this is an extension of a Recurrent Neural Network (RNN) and how the underlying network works. - Recurrent Neural Networks (RNN) are a robust and powerful type of neural network and is considered to be among the most encouraging algorithms for use of classification, due to the fact of having internal memory. RNNs are designed to recognise patterns in sequences of presented data or most suitably, time-series data, genomes, handwriting and stock market data. Although RNNs were conceptualised and invented back in the 1980s \cite{17} they've only really shown their potential in recent years, with the increase of computational power due to the level of sequencing and internal memory store to retrain. + Recurrent Neural Networks (RNN) are a robust and powerful type of neural network and is considered to be among the most encouraging algorithms for the use of classification, due to the fact of having internal memory. RNNs are designed to recognise patterns in sequences of presented data or most suitably, time-series data, genomes, handwriting and stock market data. Although RNNs were conceptualised and invented back in the 1980s \cite{17}, they've only really shown their potential in recent years, with the increase of computational power due to the level of sequencing and internal memory store to retrain. Due to this 'internal' memory loop, RNNs are able to remember data and adjust neurons based on failures and alternating parameters. The way this is accomplished, knowing how a standard neural network such as a feed-forward network, should initially be understood. \cite{18} - A standard, feed-forward neural network has a single data flow with an input layer, through hidden computational layers, to an output layer. Therefore any node in the network will never see the same data again. However, in an RNN data is cycled through a loop over the same node, thus two inputs into the perceptron. Decisions are influenced by previous data that it has previously learned from if any, which in turn affects output and the weights of the network. \cite{19} + A standard, feed-forward neural network has a single data flow with an input layer, through hidden computational layers, to an output layer. Therefore any node in the network will never see the same data again. However, in an RNN data is cycled through a loop over the same node, thus two inputs into the perception. Decisions are influenced by previous data that it has previously learned from if any, which in turn affects output and the weights of the network. \cite{19} \begin{center} \includegraphics[width=15cm,height=6cm]{images/rnn_ffn.png} @@ -348,18 +348,18 @@ \textit{Figure 3: A conceptual design of an LSTM cell bank - from Medium article by Shi Yan: Understanding LSTM and its diagrams}\cite{23} \end{center} - The network takes in three initial inputs, input of current time step, output from the previous LSTM unit if any, and the memory of the previous unit. Outputs, $H_t$ - output of current network, and $C_t$ - the memory of the current unit. \cite{23} + The network takes in three initial inputs, the input of the current time step, output from the previous LSTM unit if any, and the memory of the previous unit. Outputs, $H_t$ - output of current network, and $C_t$ - the memory of the current unit. \cite{23} - The various steps of the network decide what information is thrown away from the cell state, through use of a 'forget gate' which is influencted by the calculations of sigmod memory gates which influence how much of old and new memory is used $C_{t_-1}$, $H_{t-1}$ - and $X_t$, and merged together based upon importance. The section of the cell that controls the outflow memory $H_t$ and $C_t$ determines how much of the new memory should be used by the next LSTM unit. + The various steps of the network decide what information is thrown away from the cell state, through use of a 'forget gate' which is influenced by the calculations of sigmoid memory gates which influence how much of old and new memory is used $C_{t_-1}$, $H_{t-1}$ + and $X_t$, and merged based upon importance. The section of the cell that controls the outflow memory $H_t$ and $C_t$ determines how much of the new memory should be used by the next LSTM unit. \textit{For a more in-detailed explanation of exactly how the calculations are made see} \cite{22},\cite{23} and \cite{24}. - As mentioned in the formost section of related work the use of an LSTM network would be optimal for the given problem domain over the use of machine learning algorithms, for time-series data. As detailed above, LSTMs are widley used for time-series data forecasting due to being able to remember previous data and weights over long sequence spans\cite{22}\cite{25}. The flexability of LSTMs such as many-to-many models, useful \textit{"to predict multiple future time steps at once given all the previous inputs"} due to use of look-back windows and control of multiple 3D input parameters.\cite{25} + As mentioned in the foremost section of related work the use of an LSTM network would be optimal for the given problem domain over the use of machine learning algorithms, for time-series data. As detailed above, LSTMs are widely used for time-series data forecasting due to being able to remember previous data and weights over long sequence spans\cite{22}\cite{25}. The flexibility of LSTMs such as many-to-many models, useful \textit{"to predict multiple future time steps at once given all the previous inputs"} due to use of look-back windows and control of multiple 3D input parameters.\cite{25} \subsubsection{Keras and TensorFlow} - TensorFlow is an open-source numerical math computational library framework for dataflow differentiable programming, primarily used for machine and deep learning applications such as neural networks. TensorFlow bundles various machine learning and deep learning models and algorithms into one package for the Python language, but executes as byte code executed in C++ for performance. TensorFlow provides a range of conveniences to developers for the types of algorithms it supports such as debugging models and modifying graph operations separately instead of constructing and evaluating all at once, and compatibility to execute on CPUs or GPUs \cite{26}. However, TensorFlow's implementation and API, although provides an abstraction for development for machine and deep learning algorithms and simplifies implementation, it isn't all too friendly to programmers to use, especially new developers to the field of machine and deep learning. This is were the Keras API comes in. + TensorFlow is an open-source numerical math computational library framework for dataflow differentiable programming, primarily used for machine and deep learning applications such as neural networks. TensorFlow bundles various machine learning and deep learning models and algorithms into one package for the Python language, but executes as byte code executed in C++ for performance. TensorFlow provides a range of conveniences to developers for the types of algorithms it supports such as debugging models and modifying graph operations separately instead of constructing and evaluating all at once, and compatibility to execute on CPUs or GPUs \cite{26}. However, TensorFlow's implementation and API, although provides an abstraction for development for machine and deep learning algorithms and simplifies implementation, it isn't all too friendly to programmers to use, especially new developers to the field of machine and deep learning. - Keras is a high-level built to run atop of deep learning libraries such as Tensorflow and Theanos - another deep learning library similar to Tensorflow. It is designed to further simplify the use and application of such deep learning libraries thus making implementing a neural network and similar supported algorithms friendlier to developers working in Python. It accomplishes this by being a modular API; neural layers, cost functions, optimisers, activation functions, and regularisation schemes are all standalone features of the API that can be combined to create functional or sequential models. Due to being a high-level API for more refined and easier development of deep learning libraries it does not provide these low-level operations and algorithms; Keras relies on a back-end engine such as TensorFlow and supports a wide range of others. + Keras is a high-level built to run atop of deep learning libraries such as Tensorflow and Theanos - another deep learning library similar to Tensorflow. It is designed to further simplify the use and application of such deep learning libraries thus making implementing a neural network and similar supported algorithms friendlier to developers working in Python. It accomplishes this by being a modular API; neural layers, cost functions, optimisers, activation functions, and regularisation schemes are all standalone features of the API that can be combined to create functional or sequential models. Due to being a high-level API for a more refined and more natural development of deep learning libraries, it does not provide these low-level operations and algorithms; Keras relies on a back-end engine such as TensorFlow and supports a wide range of others. \subsubsection{Optimisers} There are three distinct optimisers used for LSTM networks; ADAgrad optimizer, RMSprop, Adam. The role of an optimiser @@ -376,13 +376,13 @@ \[v_{t+1} = \mu v_t - \alpha \delta L(\theta_t)\]\cite{28} \begin{itemize} - \item Adagrad (Adaptive Gradient): Is a method for adaptive rate learning through adaptively changing the learning parameters. This involves performing larger updates for infrequent parameters and smaller updates for frequent parameters. This algorithm fundamentally eliminates the need to manually tune the learning rate of the neural network, and is well suited with sparse data in a large scale network. \cite{28} + \item Adagrad (Adaptive Gradient): Is a method for adaptive rate learning through adaptively changing the learning parameters. This involves performing more substantial updates for infrequent parameters and smaller updates for frequent parameters. This algorithm fundamentally eliminates the need to tune the learning rate of the neural network manually and is well suited with sparse data in a large scale network. \cite{28} \[\theta_{t+1} = \theta_t + v_{t+1} \frac{\eta}{\sqrt{G_t + \epsilon}} \cdot g_t\] \begin{center} ($G_t$ is the sum of the squares of the past gradients to $\theta$) \end{center} - \item RMSProp (Root Mean Square Propagation): Aims to resolve Adagrad’s radically diminishing learning rates by using a moving average of the squared gradient. Thus utilises the magnitude of the recent gradient decsent to normalise it, and gets adjusted automatically by choosing different learning rate for each parameter. \cite{29} + \item RMSProp (Root Mean Square Propagation): Aims to resolve Adagrad’s radically diminishing learning rates by using a moving average of the squared gradient. Thus utilises the magnitude of the recent gradient descent to normalise it, and gets adjusted automatically by choosing different learning rate for each parameter. \cite{29} \[\theta_{t+1} = \theta_t - \frac{\eta}{\sqrt{(1 - \gamma) g^2_{t-1} + \gamma g_t + \epsilon}} \cdot g_t\] @@ -390,7 +390,7 @@ ($\gamma$ - decay that takes value from 0-1. $g_t$ - moving average of squared gradients) \end{center} \cite{30} - \item Adam (Adaptive Moment Estimation): Also aims to resolve Adagrad’s diminishing learning rates, by calculates the adaptive learning rate for each parameter. Being one of the most popular gradient decsent optimisation algorithms, it estimates from the 1st and 2nd moments of gradients. Adam implements the exponential moving average of the gradients to scale the learning rate of the network, and keeps an average of past gradients. \cite{31} + \item Adam (Adaptive Moment Estimation): Also aims to resolve Adagrad’s diminishing learning rates, by calculates the adaptive learning rate for each parameter. Being one of the most popular gradient descent optimisation algorithms, it estimates from the 1st and 2nd moments of gradients. Adam implements the exponential moving average of the gradients to scale the learning rate of the network and keeps an average of past gradients. \cite{31} \[m_t = \beta_1 m_{t-1} + (1 - \beta_1) g_t\] \[v_t = \beta_2 v_{t-1} + (1 - \beta_2) g^2_t\] @@ -434,7 +434,7 @@ \item $P(H\cap A)$ is the probability of the hypothesis given the occurance of evidence of the probability \end{itemize} - The naive approach assumes the features that are used in the model are independent of one another, such that, changing the value of a feature doesn't directly influence the value of the other features used in the model. When such features are independent, the Bayes algorthim can be expanded: + The naive approach assumes the features that are used in the model are independent of one another, such that, changing the value of a feature doesn't directly influence the value of the other features used in the model. When such features are independent, the Bayes algorithm can be expanded: \[P(H\cap A) = \frac{P(A\cap H) * P(H)}{P(A)} \] @@ -446,48 +446,48 @@ \[Probability \ of \ Outcome \cap Evidence = \frac{Probability \ of \ Likelihood \ of \ evidence * Prior}{Probability \ of \ Evidence} \] - The naive Bayes approach has many applications, especially for the topic of this project in classifying the probability occurance of the next price. Although it is a robust algorithm is does have its drawbacks which make it not as suitable as a neural network for the given need of this project. The naive Bayes trap is an issue that may occur due to the size of dataset that will be used. There are however other scenarios this algorithm could be used such as classification of spam data.\cite{32} + The naive Bayes approach has many applications, especially for the topic of this project in classifying the probability occurrence of the next price. Although it is a robust algorithm has its drawbacks which make it not as suitable as a neural network for the given need of this project. The naive Bayes trap is an issue that may occur due to the size of the dataset that will be used. There are however other scenarios this algorithm could be used such as classification of spam data.\cite{32} \newpage \begin{center} \section{Solution Approach}\label{solution} \end{center} - This section will outline the solution intended to solve the problem that the problem statement identifies, with justification and reference to the research conducted in the litrature review. This will lay out the development process for the project and will tools and technologies will be explained for the particular use case in this project. + This section will outline the solution intended to solve the problem that the problem statement identifies, with justification and reference to the research conducted in the literature review. This will lay out the development process for the project and will tools and technologies will be explained for the particular use case in this project. \newline \subsection{Data gathering} - This will be the part of the system that will gather price data and tweets from relevent sources, twitter and cryptocurrency exchanges. + This will be the part of the system that will gather price data and tweets from relevant sources, Twitter and cryptocurrency exchanges. \newline \textbf{Price data} \newline \newline - Historical price data can be collected in a number methods, one being that of the exchange APIs, another through a historical price tracker who creates a CSV consisting of all prior historical data. Both have their merits and reliability for granting the needed data, however, a historical tracker who has been tracking the price every hour since the start of Bitcoin would be the better option. This is due to a couple of factors, the data in some historical trackers are an average unbiased price for Bitcoin - they track the price of all or a select few exchanges and average the hourly price. Whereas if the historical data was obtained directly from an exchange this would be biased and might not represent the true price of the currency, and thus would need averaging with other hourly prices from other exchanges. By using a historical tracker all the data is unbiased and averaged and readily available and doesn't require any requests to an API or coding needed to process data. + Historical price data can be collected in a number methods, one being that of the exchange APIs, another through a historical price tracker who creates a CSV consisting of all prior historical data. Both have their merits and reliability for granting the needed data; however, a historical tracker who has been tracking the price every hour since the start of Bitcoin would be the better option. This is due to a couple of factors, the data in some historical trackers are an average unbiased price for Bitcoin - they track the price of all or a select few exchanges and average the hourly price. Whereas if the historical data was obtained directly from an exchange this would be biased and might not represent the true price of the currency, and thus would need averaging with other hourly prices from other exchanges. By using a historical tracker, all the data is unbiased and averaged and readily available and doesn't require any requests to an API or coding needed to process data. - Live price data can be collected through the same methods, a historical price tracker and an exchange API. However, this doesn't work the same way, unfortunately, a historical price tracker isn't updated as frequently as exchange APIs thus wouldn't provide on the hour accurate data. Therefore exchange APIs should be utilised in this case and multiple to provide an unbiased average for the hourly price. Three exchanges will provide an sufficient average and the exchanges most likely to be used would be the more popular exchanges such as Coinbase, Bitfinex and Gemini + Live price data can be collected through the same methods, a historical price tracker and an exchange API. However, this doesn't work the same way; unfortunately, a historical price tracker is not updated as frequently as exchange APIs thus wouldn't provide on the hour accurate data. Therefore exchange APIs will be utilised in this case and multiple to give an unbiased average for the hourly price. Three exchanges will provide a sufficient average, and the exchanges most likely to be used would be the more popular exchanges such as Coinbase, Bitfinex and Gemini. \newline \textbf{Tweets} \newline \newline - Historical tweets can be obtained through the Twitter API, and however is not a feature of the Tweepy package - \textit{not mentioned or method on official Tweepy Documentation} \cite{33}. The Twitter API, as explained in the Litrature review, allows for historical tweets to be extracted from the platform, 100 per request and a maximum of 50 requests per month. This proposes an issue with not providing enough data, where sentiment will need to be calculated per hour. Simply put, for a year of hourly price data there will be 9050 records. Therefore the equivilent will be needed for sentiment, however the sentiment will be the average the sentiment per hour of tweets. Using one request, 100 tweets per hour, per hour, 905,000 tweets will need to be extracted to provide the data needed. A solution to this issue could be to use and create multiple accounts and manually extract data from the API and merge. Another option is the pay for the data from 3rd party companies whom have access to the Enterprise API and can pull more data, 2000 per request \\cite{7}\cite{8}. Due to price for data of these 3rd parties the former could be a suitable, but more time consuming option. + Historical tweets can be obtained through the Twitter API, and however is not a feature of the Tweepy package - \textit{not mentioned or method on official Tweepy Documentation} \cite{33}. The Twitter API, as explained in the Literature review, allows for historical tweets to be extracted from the platform, 100 per request and a maximum of 50 requests per month. This proposes an issue with not providing enough data, where the sentiment will need to be calculated per hour. Simply put, for a year of hourly price data, there will be 9050 records. Therefore the equivalent will be required for sentiment; however, the sentiment will be the average the sentiment per hour of tweets. Using a single request with 100 tweets per hour, per hour; 905,000 tweets will need to be extracted to provide the data required. A solution to this issue could be to use and create multiple accounts and manually extract data from the API and merge. Another option is the pay for the data from 3rd party companies who have access to the Enterprise API and can pull more data, 2000 per request \\cite{7}\cite{8}. Due to the price for data of these 3rd parties the former could be a suitable, but more time-consuming option. - Live tweets can be collected by two methods from Twitter, from the Twitter API and using Twitter Python package such as Tweepy, detailed in the Literature review. Additionally, the limitations of the Twitter API are also discussed in the review which states how the Twitter API has a tiering system: Standard, Premium and Enterprise. Each tier has different levels of access to the API and can extract a different amount of data from the platform. Thus concluding the section in the Literature review, the Twitter API will not be used for the extraction and streaming of live tweets due to it being restricted to Enterprise users. Therefore, Tweepy will be used to set up a looping authenticated streaming solution with the Twitter API which will allow the access of current recurring data. + Live tweets can be collected by two methods from Twitter, from the Twitter API and using Twitter Python package such as Tweepy, detailed in the Literature review. Additionally, the limitations of the Twitter API are also discussed in the literature review which states how the Twitter API has a tiering system: Standard, Premium and Enterprise. Each tier has different levels of access to the API and can extract varying amounts of data from the platform. Thus concluding the section in the Literature review, the Twitter API will not be used for the extraction and streaming of live tweets due to it being restricted to Enterprise users. Therefore, Tweepy will be used to set up a looping authenticated streaming solution with the Twitter API which will allow the access of current recurring data. Natural language pre-processing will be apart of most systems in this project. Techniques such as tokenisation, stemming, stopword removal and character filtering will be prevalent, as these will be used to remove unwanted data and to sanitise the data for classification. \subsection{Data pre-processing} Natural language pre-processing will be apart of most systems in this project. Techniques such as tokenisation, stemming, stopword removal and character filtering will be prevalent, as these will be used to remove unwanted data and to sanitise the data for classification. \subsection{Spam Filtering} - This part of the system will aim to detect whether or not the steamed and/or the historical tweet is spam - unwanted tweets that serve no purpose in determining opinion of the public. These types of tweets can be from advertisement - usually labeled with \textit{\#Airdrop} and can contain \textit{"tickets here" and "Token Sale"}, to job advertisments - usually containing word such as \textit{Firm, hire, hiring, jobs and careers}. It is important to filter out and remove such data from the network as these can be seen as outliers of the true needed data and will skew predictions will invalid sentiment. + This part of the system will aim to detect whether or not the streamed data or the historical data is spam - unwanted tweets that serve no purpose in determining the opinion of the public. These types of tweets can be from advertisement - usually labelled with \textit{\#Airdrop} and can contain \textit{"tickets here" and "Token Sale"}, to job advertisements - usually containing word such as \textit{Firm, hire, hiring, jobs and careers}. It is essential to filter out and remove such data from the network as these can be seen as outliers of the true data and will skew predictions will invalid sentiment. - The spam filter should use a probability-based algorithms such as Naive Bayes, other algorithms such as ... could be used, but due to this being a probability related problem using an algorithm such as Naive Bayes would be more suitable. This classifier should be trained on a hand created dataset containing both spam and ham (\textit{wanted data}) tweets, and should not be exclusive to either category. + The spam filter will use a probability-based algorithm such as Naive Bayes, other algorithms such as Random Forest could be used, but due to this being a probability related problem using an algorithm such as Naive Bayes would be more suitable. This classifier will be trained on a hand created dataset containing both spam and ham (\textit{wanted data}) tweets, and should not be exclusive to either category. \subsection{Language Detection} - Pior to performing any kind of natural languge pre-processing and spam filtering non-English tweets will need to be avoided. This can be introduced through various language detection filtering using techniques such as ngrams alongside other natural language pre-processing techniques to filter out non-English characters. Fortunatly both Tweepy and the Twitter API have methods for specifying the desired language to recieve tweets in - \textit{filter=['en']} for the Tweepy streaming method and \textit{query=\{...,language='en',...\}} on the JSON parameters for the Twitter API. This does provide a simply means of filtering out non-English tweets, but this only filters based on region and user settings which indicate the users desired language. Thus if a user has their region set to \textit{'en'} or has their desired language set also as \textit{'en'} the tweet will be classified as English but may contain non-English characters. + Before performing any natural language pre-processing and spam filtering, non-English tweets will need to be reduced. This can be introduced through various language detection filtering using techniques such as ngrams alongside other natural language pre-processing techniques to filter out non-English characters. Fortunately, both Tweepy and the Twitter API have methods for specifying the desired language to receive tweets in - \textit{filter=['en']} for the Tweepy streaming method and \textit{query=\{...,language='en',...\}} on the JSON parameters for the Twitter API. This provides a simple means of filtering out non-English tweets, but this only filters based on region and user settings which indicate the users desired language. Thus if a user has their region set to \textit{'en'} or has their desired language set also as \textit{'en'} the tweet will be classified as English but may contain non-English characters. - As being the case a suitable language detection system will be implemented to identify any tweets that contain non-English character make it past the inital API filters, and will drop the tweets if it contains more non-English characters. If, however, the majority of the text is English but contains some non-English characters, these will be removed from the tweet. + As is the case, a suitable language detection system will be implemented to identify any tweets that contain non-English characters. Some tweet will regrettably make it past the initial API filters; thus such a system will be implemented that will drop the tweets if it predominately contains non-English characters. If however, the majority of the text in English but includes some non-English characters, these will be removed from the tweet. \subsection{Sentiment Analysis} As mentioned in the Litrature review, the VADER sentiment analysis performs exceptionally well on the social media domain when compared to idividual human rates and 10 other highly regarded sentiment analysers, stated in the results section of the paper \textit{VADER: A Parsimonious Rule-based Model for Sentiment Analysis of Social Media Text} \cite{12}. \newline Extraction of results from paper \cite{12}: @@ -510,42 +510,51 @@ \textbf{Analysis of Social Media Text (4,200 Tweets)}\cite{12} \end{center} - Due to the suitability for the given domain of social media and with the customisability, due to VADER's lexicon-dictionary based approach, makes this sentiment analyser most suitable for use in this project. This analyser will be utilised as the sentiment analyser of this project due to its feature set and need for little data pre-processing prior to polarity classification of the provided text. \cite{11} \textit{"extract"}. + Due to the suitability for the given domain of social media and with the customisability, due to VADER's lexicon-dictionary based approach, makes this sentiment analyser most suitable for use in this project. This analyser will be utilised as the sentiment analyser of this project due to its feature set and need for little data pre-processing before polarity classification of the provided text. \cite{11} \textit{"extract ..."}. - This will be an itermediate system between the neural network and the data collection pre-processing system, as the later will provide the cleaned processed data for analysis and the former to feed in the classified polarity of each tweet alongside price data for model learning. + This will be an intermediate system between the neural network and the data collection pre-processing system, as the later will provide the cleaned processed data for analysis and the former to feed in the classified polarity of each tweet alongside price data for model learning. \subsection{Neural Network} - The \textit{Neural Network} section in the litrature review details how Recurrent Neural networks work alongside how an Long-short term memory networks build upon and overcome limitations and known issues with a standard RNN network. A recurrent neural network is the focus of this project, this is due to: + The \textit{Neural Network} section in the literature review details how Recurrent Neural networks work alongside how a Long-short term memory networks build upon and overcome limitations and known issues with a standard RNN network. A recurrent neural network is the focus of this project, and this is due to: \begin{itemize} - \item Nature of an RNN - Allows for backpropagation to find partial derivatives of the error with respect to the weights after an output has occured, to tweak the current weights of the LSTM cell. In short, allows the tweaking of weights of the network based on previous seen data by looping the same node thus influencing decisions made on current data based on old weights and errors from previous. - \item Nature of an LSTM over RNN - LSTMs are extensions of RNNs \cite{22} that were designed to avoid long-term dependancy problems such as exploding and vanishing gradients. Weights are not only just reused but are stored in memory and are propogated through the network. - \item Lack of use for the projects purose - Other papers tend to focus on machine learning techniques, other neural networks such as Multi-layer Perceptron (MPL) and standard Recurrent Neural Networks, with use of time-series data. Especially with use of a standard RNN, not overcomming its common issues with gradient decsent. Stated in related research section of the litrature review, \cite{5} - \textit{"using the MLP classifier (a.k.a neural networks) showed better results than logistical regression and random forest trained models"} - \item Pior use for time-series data and data forecasting - Although RNN LSTM networks have been used for the prediction of Bitcoin price there are a few papers on this \cite{25}. Regardless, LSTMs have been notably used with use for time-series data forecasting due to being able to remember previous data and weights over long sequence spans \cite{22} - \textit{""}, \cite{25} - \textit{""}. + \item Nature of an RNN - Allows for backpropagation to find partial derivatives of the error with respect to the weights after an output has occurred, to tweak the current weights of the LSTM cell. In short, allows the tweaking of weights of the network based on previously seen data by looping the same node thus influencing decisions made on current data based on old weights and errors from previous. + \item Nature of an LSTM over RNN - LSTMs are extensions of RNNs \cite{22} that were designed to avoid long-term dependency problems such as exploding and vanishing gradients. Weights are not only just reused but are stored in memory and are propagated through the network. + \item Lack of use for the project's purpose - Other papers tend to focus on machine learning techniques, other neural networks such as Multi-layer Perceptron (MPL) and standard Recurrent Neural Networks, with use of time-series data. Especially with the use of a standard RNN, not overcoming its common issues with gradient descent. Stated in related research section of the literature review, \cite{5} - \textit{"using the MLP classifier (a.k.a neural networks) showed better results than logistical regression and random forest trained models"} + \item Prior use for time-series data and data forecasting - Although RNN LSTM networks have been used for the prediction of Bitcoin price there are a few papers on this \cite{25}. Regardless, LSTMs have been notably used with use for time-series data forecasting due to being able to remember previous data and weights over long sequence spans \cite{22} - \textit{""}, \cite{25} - \textit{""}. \end{itemize} - Therefore, a recurrent long-short term memory neural network will be used for this for this project to predict the next hour interval of Bitcoin price based on previous historical prices and hourly sentiment. This system will read in historical data, both price and sentiment - depending on the network for prediction with and without sentiment, this data will be merged, split and used to trained and test the network model for use for forecasting prices. The relative sizes for the training and test data can be decided upon system creation but the standard sizing is 75:25 respectivly. + Therefore, a recurrent long-short-term memory neural network will be used for this project to predict the next hour interval of Bitcoin price based on previous historical prices and hourly sentiment. This system will read in historical data, both price and sentiment - depending on the network for prediction with and without sentiment, this data will be merged, split and used to trained and test the network model for use for forecasting prices. The relative sizes for the training and test data can be decided upon system creation, but the standard sizing for training neural networks is 75:25 respectively. - Tensorflow will also be used for the backend implementation and the Kera API use upon it to make development more straight forward. There are other tools that are comparable to TensorFlow that are also supported by Keras, such as \textit{"TABLE COMPARING TOOLS"} ... + Tensorflow will be used for the network implementation, and the Keras API use upon it to make development more straight forward. Other tools are comparable to TensorFlow that are also supported by Keras. + \begin{table}[ht] + \centering + \resizebox{\textwidth}{!}{\begin{tabular}{l|c|c} + \textbf{Framework} & \textbf{Pros} & \textbf{Cons}\\ + \hline + \multirow{6}{*}{TensorFlow} & Supports reinforcement learning and other algorithms & Doesn’t support matrix operations \\ & Offers computational graph abstraction & Doesn't have pertained models \\ & Faster compile time than Theano & Drops to Python to load each new training batch \\ & Data and model parallelism & Doesn't support dynamic typing on large scale projects \\ & Can be deployed over multiple CPUs and GPUs & \\ + \hline + \multirow{4}{*}{Theano} & Computational Graph Abstraction & Is low-level \\ & Has multiple high-level wrappers similar to Keras & Can only be deployed to a single GPU \\ & & Much slower compile times on large models than competition \\ & & Unhelpful and vague error messages \\ + \hline + \multirow{3}{*}{Pytorch} & Graph definition is more imperative and dynamic than other frameworks & Not as widley adopted as TensorFlow \\ & Graph computation defined at runtime, allowing standard popular IDEs to support it & Visualisation is not as robust as TensorBoard \\ & Natively support common python deployment frameworks such as Flask & Not as deployable as TensorFlow, doesn't supper gRPC \\ & & \\ + \end{tabular}} + + \textbf{Comparison between TensorFlow, Theano and Pytorch}\cite{34} + \end{table} \subsection{Price Forecasting} - This part of the system will be responsible for prediction the next time-step of Bitcoin's price for the next hour based on past data. It will use the trained model from the neural network to predict the next hour price when given live hourly data, price and sentiment. I will also have a look back of 5 which will allow it to see historical data to aid in the predictions. This will ocurr on the hour every hour when new data is recieved and processed, this data should also be merged and the split into training and testing data, the sizing can be decided upon system creation but the standard sizing is 75:25, training and testing respectivly. + This part of the system will be responsible for prediction the next time-step of Bitcoin's price for the next hour based on past data. It will use the trained model from the neural network to predict the future hour price when given live hourly data, price and sentiment. The system will also have a look back of 5 which will allow it to see historical data to aid in the predictions. This will occur on the hour every hour when new data is received and processed, this data will also be merged and the split into training and testing data. The sizing can be decided upon system creation, but the standard sizing for training is 75:25, training and testing respectively. \subsection{Frontend Application} - The frontend appliction will display the predicted data to the stakeholders and users of the system, along with charting True hourly prices against Predicted, for both with and without sentiment embedded in the predictions. The interface will display this data in both tabular and chart form to provide variety to the user. Performance metrics will also be disayed at the bottom of the appliction to show the accuracy of the model. Due to this project focusing around the backend, how the predictions are made and the accuracy of the model, the interface will be somewhat of a second thought. It will aim to display the information in a clear and concise manner which will start to solve the problem of providing a system to the public to aid in investment decisions, the design will not be complex but more basic and functional. Therefore a basic webpage coded in HTML with Jquery to plot data, and Ajax requests to obtain and load data, will be sufficient. - - %\includegraphics[width=18cm,height=8cm]{images/interface_design.png} - \begin{center} - \textit{Figure 3: Initial Interface design} - \end{center} + The frontend application will display the predicted data to the stakeholders and users of the system, along with charting True hourly prices against Predicted, for both with and without sentiment embedded in the predictions. The interface will display this data in both tabular and chart form to provide variety to the user. Performance metrics will also be displayed at the bottom of the application to show the accuracy of the model. Due to this project focusing around the backend, how the predictions are made and the accuracy of the model, the interface will be somewhat of a second thought. It will aim to display the information in a clear and concise manner which will start to solve the problem of providing a system to the public to aid in investment decisions. The design will not be complicated but more basic and functional. Therefore a basic webpage coded in HTML with Jquery to plot data, and Ajax requests to obtain and load data, will be sufficient. \subsection{With reference to Initial PID} - Both the problem and solution has changed considerably from the initial project initiation document (PID), which outlines the inital ideas, objectives and specification for the project. The reason for this was due to a change in direction which was caused by a number of factors; one being a change in passion after intial research into machine learning techniques and neural networks, instead of creating an application that just performed sentiment analysis the direction turned towards how this could be used to predict future prices. This change does still loosly keeps in-line with the intial idea of wanting to create a platform that will aid in investor decision making, but takes it a step further by directly giving them predictions on market direction price as a basis for these decisions rather than just idenifying opinion direction of the market. - Another point was simplicity of the inital idea, which consisted of focusing more work on the design of the frontend appliction to display opinion data and general price data on a range of cryptocurrencies which will simply by just consuming exchange APIs. Both the developer and project supervisor came to the conclusion that this inital idea was too simple and a more complex approach needed to be formed. - The intial PID did however give an initial basis to base ideas and inital research from and was the begining drive of this project. + Both the problem and solution have changed considerably from the original project initiation document (PID), which outlines the initial ideas, objectives and specification for the project. The reason for this was due to a change in direction which was caused by a number of factors; one being a change in passion after initial research into machine learning techniques and neural networks, instead of creating an application that just performed sentiment analysis the direction turned towards how this could be used to predict future prices. This change does still loosely keeps in-line with the initial idea of wanting to create a platform that will aid in investor decision making but takes it a step further by directly giving them predictions on market direction price as a basis for these decisions rather than just identifying opinion direction of the market. + Another point was the simplicity of the initial idea, which consisted of focusing more work on the design of the frontend application to display opinion data and general price data on a range of cryptocurrencies which will only by consuming exchange APIs. Both the developer and project supervisor concluded that this initial idea was too simple and a more sophisticated approach needed forming. + The initial PID did, however, give an initial basis to base ideas and initial research from and was the beginning drive of this project. \subsection{Solution Summary}\label{summary} - The overall solution, with reference to the problem statement, is to create a system mainly consisting of; a frontend application that will display plotting, predicted and true, performance metric data to the user as a clear and concise form. A backend system that'll be behind the price forecasting, which will consist of various subsystem responsible for data collection, filtering, data pre-processing, sentiment analysis, network training, validation and training and future price predictions. Each stage will consist of revelent tools and techniques for performing their required task. + The overall solution, concerning the problem statement, is to create a system mainly consisting of; a frontend application that will display plotting, predicted and true, performance metric data to the user as a clear and concise form. The backend system behind the price forecasting will consist of various subsystem responsible for data collection, filtering, data pre-processing, sentiment analysis, network training, validation and training and future price predictions. Each stage will consist of relevant tools and techniques for performing their required task. %The tools and techniques that will be used for this project are as follows, with relation to the relevent part of the system. @@ -564,7 +573,7 @@ \includegraphics[width=18cm,height=8cm]{images/Generic_Flow.png} \begin{center} - \textit{Figure 4: Basic Dataflow diagram of systems in the project and how data could possibly flow} + \textit{Figure 3: Basic Dataflow diagram of systems in the project and how data could possibly flow} \end{center} \newpage @@ -573,73 +582,73 @@ \section{System Design}\label{Design} \end{center} \subsection{Dataflow Designs} - This section will describe and outline how the system will be formed and will work with each component, a good way of displaying this is as a dataflow diagram. A dataflow is a way of representing the flow of data through a process or system, as a result it also provides information about how inputs and outputs of each component works and how they're connected to other components. It can also give either broad or in-depth overview of the specific workings of each component through how the data is processed and manipulated. + This section will describe and outline how the system will be formed and will work with each component; a useful way of displaying this is as a dataflow diagram. A dataflow is a way of representing the flow of data through a process or system; as a result, it also provides information about how inputs and outputs of each component work and how they function with other components. It can also give either broad or in-depth overview of the specific workings of each component through how the data is processed and manipulated. \newline \textbf{Dataflow overview of entire system:} \begin{center} \includegraphics[width=18cm,height=8cm]{images/Dataflow.png} - \textit{Figure 5: Overall Dataflow diagram of the entire system} + \textit{Figure 4: Overall Dataflow diagram of the entire system} \end{center} - This dataflow diagram shows the overall concept of how the data is intended to flow through the system, from being processed and manipulated through each components and what the outputs are of each. Due to the size this will be broken up and individually explained. + This dataflow diagram shows the overall concept of how the data is intended to flow through the system, from being processed and manipulated through each component and what the outputs are of each. Due to the size, this will be broken up and individually explained. \newpage \textbf{Data collector} \begin{center} \includegraphics[width=15cm,height=8cm]{images/Data_Collector.png} - \textit{Figure 6: Data collector Dataflow diagram} + \textit{Figure 5: Data collector Dataflow diagram} \end{center} This dataflow diagram shows the part of the system responsible for the collection and processing of both historical data. This is split into three parts: Price collector, Tweet collector and tweet normalisation and natural language pre-processing. \begin{itemize} \item Price Collector - Processes two forms of data, Historical and Live price data. - \subitem Historical data is extrapolated from three CSVs that contain the historical price every hour for the past year, from a historical price tracker. At this point in the project it was identified that historical price trackers do not average price data from exchanges as previously identified, therefore this data will need to be merged and averaged to create the unbiased hourly price needed. + \subitem Historical data is extrapolated from three CSVs that contain the historical price every hour for the past year, from a historical price tracker. At this point in the project, it was identified that historical price trackers do not average the price data from exchanges as previously identified; therefore this data will need to be merged and averaged to create the unbiased hourly price needed. \subitem Live data is extracted directly from the three exchanges APIs shown through REST endpoint requests. - \subitem Data from both, as separate processes independant from one another, are averaged by extracting the \textit{High}, \textit{Mid} and \textit{Low} hourly prices. This averaged price per hour for each exchange are then averaged together to obtain an unbiased hourly average. This is then saved to a CSV of historical or live prices respectivly. The difference in the flow of data is that of Live prices, in which the process is looped every hour to extract the new hourly prices. + \subitem Data from both, as separate processes independent from one another, are averaged by extracting the \textit{High}, \textit{Mid} and \textit{Low} hourly prices. This averaged price per hour for each exchange are then averaged together to obtain an unbiased hourly average. The price is then saved to a CSV of historical or live prices respectively. The difference in the flow of data is that of Live prices, in which the process is looped every hour to extract the new hourly prices. \item Tweet Collector - Streams tweets from Twitter using Tweepy, historical tweets are manually collected directly from the Twitter API. Both are fed through the normalisation and data pre-processing stage. - \item Data pre-processing - This involves cleaning the intial data by removing line breaks and new lines that occur in the data, removal of special characters that are standard in tweets (\textit{'\#','\@' and urls}). This is then fed into a language detection system which tokenises and compares stopwords in text to NLTK package supported languages. Depending on whether the text is idendified as majoritly English or not determines whether or not the tweet is dropped and not used in the network. If the majority is in English, non-English characters are removed as these can still be present in the text. + \item Data pre-processing - This involves cleaning the initial data by removing line breaks and new lines that occur in the data, removal of special characters that are standard in tweets (\textit{'\#','\@' and urls}). The data is then fed into a language detection system which tokenises and compares stopwords in the text to NLTK package supported languages. Depending on whether the text is identified as being predominately English or not determines whether or not the tweet is dropped and not used in the network. If the majority is in English, non-English characters are removed as these can still be present in the text. \end{itemize} \textbf{Analysis Engine} \begin{center} \includegraphics[width=17cm,height=8cm]{images/Analysis_Engine.png} - \textit{Figure 7: Analysis Engine Dataflow diagram} + \textit{Figure 6: Analysis Engine Dataflow diagram} \end{center} - This dataflow diagram shows the part of the system that is responsible for training a spam filter, creating the model that'll be used to identify if the tweets from the data collector are unwanted - spam. This system is also responsible for assigning the polarity classifiction to the tweet through sentiment analysis conducted by the VADER package \cite{12}. + This dataflow diagram shows the part of the system that is responsible for training a spam filter, creating the model that'll be used to identify if the tweets from the data collector are unwanted - spam. This system is also responsible for assigning the polarity classification to the tweet through sentiment analysis conducted by the VADER package \cite{12}. \begin{itemize} - \item Spam filter training - The inital step in this system is to train the Naive Bayes Classifier using the pre-labeled spam dataset which contains an unbiased amount of either spam or ham tweets with their respective labels. + \item Spam filter training - The initial step in this system is to train the Naive Bayes Classifier using the pre-labelled spam dataset which contains an unbiased amount of either spam or ham tweets with their respective labels. \subitem This data is split into two samples, training and test sets 75:25 respectively and the Naive Bayes classifier trained and validated against these datasets after pre-processing of the data occurs on the data to prepare it. \item Data pre-processing - The tweets from both training and testing the filter and from live and historical tweets are processed through this section. \subitem This section of the system is primarily used to process the tweets for the filter to classify the data and doesn't directly modify the live and historical tweets. The data is processed through various natural language processing techniques such as; Tokenisation, Ngram generation, stopword removal and stemming. - \item Classifier Modelling and Model creation - Once the data is pre-processed the data is classified and the prediction model created, which later used to classify the historical and live tweets. - \item Sentiment Analysis (VADER) - On a separate route from the spam filter training, using the historical and live tweets, the sentiment analysier VADER performs analysis on the tweets and assigns a polarity classification to each text (\textit{Negative, Neutral, Positive} and calculates the compound score which is the difference between the negative and positive scores \textit{compound}). - \item Storage - The polarity classification and tweets are saved to their respective CSV files for historical and live data. + \item Classifier Modelling and Model creation - Once the data is pre-processed, the data is classified, and the prediction model created, which later used to classify the historical and live tweets. + \item Sentiment Analysis (VADER) - On a separate route from the spam filter training, using the past and live tweets, the sentiment analyser VADER performs analysis on the tweets and assigns a polarity classification to each text (\textit{Negative, Neutral, Positive} and calculates the compound score which is the difference between the negative and positive ratings \textit{compound}). + \item Storage - The polarity classification and tweets are then saved to their relevant CSV files for historical and live data. \end{itemize} \textbf{Neural Network} \begin{center} \includegraphics[width=17cm,height=12cm]{images/Neural_Network.png} - \textit{Figure 8: Neural Network layout Dataflow diagram} + \textit{Figure 7: Neural Network layout Dataflow diagram} \end{center} - The dataflow diagram in \textit{figure 8} shows the part of the system that is responsible for training and creating the neural network model. The dataflow diagram show how this will be trained and the layers of a possible solution to the network, which shows 4 layers which may not be the solution that will be implemented but are there to show a representation of an amount of layer that could be implemented. + The dataflow diagram in \textit{figure 7} shows the part of the system that is responsible for training and creating the neural network model. The dataflow diagram shows how the network will be trained, and the layers of a possible solution to the network. The model shows four layers which may not be the solution that will be implemented but is there to show a representation of a number of layers that could be applied. \begin{itemize} - \item Merging of Datasets - Data from both historical datasets are merged to create one dataset with mapped price and sentiment for each hour. *This is a specific process that is different with the system that does not include sentiment for predictions, the merge process doesn't occur in that system/model. + \item Merging of Datasets - Data from both historical datasets are merged to create one dataset with mapped price and sentiment for each hour. *This is a specific process that is different from the system that does not include sentiment for predictions, the merge process doesn't occur in that system/model. \item Training and Testing - Data is split into two samples of training and testing, 75:25 respectively. **This also doesn't occur in the system that doesn't model with the sentiment. - \item Training network - The training sets, X and Y coords are used to train the network. - \item Testing network - The testing sets, X and Y coords of 25\% of the initial data are used to test the validation and accuracy of predictions as these contain the true data of what the predictions should be. + \item Training network - The training sets, X and Y coordinates are used to train the network. + \item Testing network - The testing sets, X and Y coordinates of 25\% of the initial data are used to test the validation and accuracy of predictions as these contain the true data of what the predictions should be. \item Outputs - Accuracy Statistics, true price data and predicted next hour prices are outputted to respective files for use on the front-end application. The model is then later used for hourly forecasting. \end{itemize} \textbf{Future Price Forecasting} \begin{center} \includegraphics[width=18cm,height=8cm]{images/Future_Predictions.png} - \textit{Figure 9: Price Forecasting Dataflow diagram} + \textit{Figure 8: Price Forecasting Dataflow diagram} \end{center} - The dataflow diagram in \textit{figure 9} shows how the forecasting system would be implemented. This dataflow shows how it will read live data of both sentiment and price data, merge, split and conduct regression using the trained neural network model to predict the next hour price. + The dataflow diagram in \textit{figure 8} shows how the forecasting system would be implemented. This dataflow shows how it will read live data of both sentiment and price data, merge, split and conduct regression using the trained neural network model to predict the next hour price. \begin{itemize} - \item Data merging - (Doesn't occur with the system that doesn't include sentiment in price predictions) Data is merged from both historical and live data up to 5 iterations. This is due to after the initial hour there will only be one record of price and sentiment data, in which not prediction will be made from this as there isn't sufficient amount of data. + \item Data merging - (Doesn't occur with the system that doesn't include sentiment in price predictions). Data is consolidated from both historical and live data up to 5 iterations. This is due to after the initial hour there will only be a singular record of price and sentiment data, in which no prediction could be made from this as there isn't a sufficient amount of data. \item Prediction - This data is then fitted to the neural network model and predictions for the next time-step hour are made. \item Hour Loop - This will then proceed to loop every hour to make the hourly predictions. Historical price data will cease to be used when there are 5 or more live price records. - \item Outputs - Accuracy Statistics, true price data and predicted next hour prices are ouptutted to respective files for use on the front-end application for charting. + \item Outputs - Accuracy Statistics, true price data and predicted next hour prices are outputted to respective files for use on the front-end application for charting. \end{itemize} \newpage @@ -647,20 +656,30 @@ \begin{center} \includegraphics[width=10cm,height=9cm]{images/Frontend_Application.png} \newline - \textit{Figure 10: Front-end Application Dataflow diagram} + \textit{Figure 9: Front-end Application Dataflow diagram} \end{center} The above dataflow diagram shows the data flow for the front-end application and how the data is read into the system from the data files generated by the backend application (Neural network). \begin{itemize} - \item Ajax Requests - These are api file requests for files hosted on the server in which the system is running on. This loads the data files into the application for use. - \item CSS Styling - Contains design styling for page and charts, loaded upon loading of webpage. - \item Charting and Tables - Accesses the loaded data from the Ajax requests and plots the data. Prediction data, only with sentiment and prices are plotted into a table. There will be separate charts and tables displaying the data from the backend that hasn't used sentiment in predictions to aid in establishing a correlation between sentiment and price and whether it affects the hourly price (Aiming to solve problem statement) - \item Stakeholders - There will be the four stakeholder, outline in the problem articulation section, that would be the primary users of this application. + \item Ajax Requests - These are API file requests for files hosted on the server in which the system is running on. This loads the data files into the application for use. + \item CSS Styling - Contains design styling for page and charts, loaded upon loading of a webpage. + \item Charting and Tables - Accesses the loaded data from the Ajax requests and plots the data. Prediction data, only with sentiment and prices are plotted into a table. There will be separate charts and tables displaying the data from the backend that hasn't used sentiment in predictions to aid in establishing a correlation between sentiment and price and whether it affects the hourly price (Aiming to solve the problem statement) + \item Stakeholders - There will be the four stakeholders, outlined in the problem articulation section, that would be the primary users of this application. \end{itemize} \newpage \subsection{UML Component Design} \subsection{Interface Design} + + \begin{figure}[hbt!] + \centering + \includegraphics[width=8cm,height=13cm]{images/interface_design.png} + \end{figure} + \begin{center} + \textit{Figure 10: Interface design} + \end{center} + Figure 10 above shows the basic idea of the interface design that will be presented to the stakeholders and aims to be the interface that these stakeholders will use to aid in their market decisions of Bitcoin. The interface, although simplistic, provides all the necassary information that any of these stakeholders would need, it also provides information to allow visual comparision on how sentiment affects the hourly price of Bitcoin, represented as the two charts. + \newpage \begin{center} \section{Implementation}\label{implementation} diff --git a/document.toc b/document.toc index b8a2a3b..34d6ac4 100644 --- a/document.toc +++ b/document.toc @@ -62,76 +62,76 @@ \defcounter {refsection}{0}\relax \contentsline {subsection}{Language Detection}{24}{section*.32} \defcounter {refsection}{0}\relax -\contentsline {subsection}{Sentiment Analysis}{24}{section*.33} +\contentsline {subsection}{Sentiment Analysis}{25}{section*.33} \defcounter {refsection}{0}\relax \contentsline {subsection}{Neural Network}{25}{section*.34} \defcounter {refsection}{0}\relax -\contentsline {subsection}{Price Forecasting}{26}{section*.35} +\contentsline {subsection}{Price Forecasting}{26}{section*.36} \defcounter {refsection}{0}\relax -\contentsline {subsection}{Frontend Application}{26}{section*.36} +\contentsline {subsection}{Frontend Application}{26}{section*.37} \defcounter {refsection}{0}\relax -\contentsline {subsection}{With reference to Initial PID}{26}{section*.37} +\contentsline {subsection}{With reference to Initial PID}{27}{section*.38} \defcounter {refsection}{0}\relax -\contentsline {subsection}{Solution Summary}{27}{section*.38} +\contentsline {subsection}{Solution Summary}{27}{section*.39} \defcounter {refsection}{0}\relax -\contentsline {subsection}{Data flow Overview}{27}{section*.39} +\contentsline {subsection}{Data flow Overview}{27}{section*.40} \defcounter {refsection}{0}\relax -\contentsline {section}{System Design}{28}{section*.40} +\contentsline {section}{System Design}{29}{section*.41} \defcounter {refsection}{0}\relax -\contentsline {subsection}{Dataflow Designs}{28}{section*.41} +\contentsline {subsection}{Dataflow Designs}{29}{section*.42} \defcounter {refsection}{0}\relax -\contentsline {subsection}{UML Component Design}{34}{section*.42} +\contentsline {subsection}{UML Component Design}{35}{section*.43} \defcounter {refsection}{0}\relax -\contentsline {subsection}{Interface Design}{34}{section*.43} +\contentsline {subsection}{Interface Design}{35}{section*.44} \defcounter {refsection}{0}\relax -\contentsline {section}{Implementation}{35}{section*.44} +\contentsline {section}{Implementation}{36}{section*.46} \defcounter {refsection}{0}\relax -\contentsline {subsection}{Data collection}{35}{section*.45} +\contentsline {subsection}{Data collection}{36}{section*.47} \defcounter {refsection}{0}\relax -\contentsline {subsubsection}{Price Time-series Data}{35}{section*.46} +\contentsline {subsubsection}{Price Time-series Data}{36}{section*.48} \defcounter {refsection}{0}\relax -\contentsline {subsection}{Data processing}{35}{section*.47} +\contentsline {subsection}{Data processing}{36}{section*.49} \defcounter {refsection}{0}\relax -\contentsline {subsubsection}{Preprocessing}{35}{section*.48} +\contentsline {subsubsection}{Preprocessing}{36}{section*.50} \defcounter {refsection}{0}\relax -\contentsline {paragraph}{Tweet Filtering}{35}{section*.49} +\contentsline {paragraph}{Tweet Filtering}{36}{section*.51} \defcounter {refsection}{0}\relax -\contentsline {paragraph}{Text Cleaning}{35}{section*.50} +\contentsline {paragraph}{Text Cleaning}{36}{section*.52} \defcounter {refsection}{0}\relax -\contentsline {paragraph}{Ngram based Language detection filtering}{35}{section*.51} +\contentsline {paragraph}{Ngram based Language detection filtering}{36}{section*.53} \defcounter {refsection}{0}\relax -\contentsline {subsubsection}{Spam Filtering}{35}{section*.52} +\contentsline {subsubsection}{Spam Filtering}{36}{section*.54} \defcounter {refsection}{0}\relax -\contentsline {paragraph}{Tweet Processing}{35}{section*.53} +\contentsline {paragraph}{Tweet Processing}{36}{section*.55} \defcounter {refsection}{0}\relax -\contentsline {paragraph}{Naive Bayes model}{35}{section*.54} +\contentsline {paragraph}{Naive Bayes model}{36}{section*.56} \defcounter {refsection}{0}\relax -\contentsline {subsection}{Sentiment Analysis}{35}{section*.55} +\contentsline {subsection}{Sentiment Analysis}{36}{section*.57} \defcounter {refsection}{0}\relax -\contentsline {subsubsection}{VADER}{35}{section*.56} +\contentsline {subsubsection}{VADER}{36}{section*.58} \defcounter {refsection}{0}\relax -\contentsline {subsection}{Recurrent Neural Network - LSTM}{35}{section*.57} +\contentsline {subsection}{Recurrent Neural Network - LSTM}{36}{section*.59} \defcounter {refsection}{0}\relax -\contentsline {subsubsection}{Training and Testing Model}{35}{section*.58} +\contentsline {subsubsection}{Training and Testing Model}{36}{section*.60} \defcounter {refsection}{0}\relax -\contentsline {subsubsection}{Scoring and Validation}{35}{section*.59} +\contentsline {subsubsection}{Scoring and Validation}{36}{section*.61} \defcounter {refsection}{0}\relax -\contentsline {subsubsection}{Future Prediction Forecasting}{35}{section*.60} +\contentsline {subsubsection}{Future Prediction Forecasting}{36}{section*.62} \defcounter {refsection}{0}\relax -\contentsline {section}{Testing: Verification and Reflection}{36}{section*.61} +\contentsline {section}{Testing: Verification and Reflection}{37}{section*.63} \defcounter {refsection}{0}\relax -\contentsline {section}{Discussion: Contribution and Reflection}{37}{section*.62} +\contentsline {section}{Discussion: Contribution and Reflection}{38}{section*.64} \defcounter {refsection}{0}\relax -\contentsline {subsection}{Limitations}{37}{section*.63} +\contentsline {subsection}{Limitations}{38}{section*.65} \defcounter {refsection}{0}\relax -\contentsline {section}{Conclusion and Future Improvements}{38}{section*.64} +\contentsline {section}{Conclusion and Future Improvements}{39}{section*.66} \defcounter {refsection}{0}\relax -\contentsline {subsection}{Conclusion}{38}{section*.65} +\contentsline {subsection}{Conclusion}{39}{section*.67} \defcounter {refsection}{0}\relax -\contentsline {subsection}{Future Improvements}{38}{section*.66} +\contentsline {subsection}{Future Improvements}{39}{section*.68} \defcounter {refsection}{0}\relax -\contentsline {section}{Appendices}{42}{section*.68} +\contentsline {section}{Appendices}{43}{section*.70} \defcounter {refsection}{0}\relax -\contentsline {subsection}{Appendix A - Project Initiation Document}{42}{section*.69} +\contentsline {subsection}{Appendix A - Project Initiation Document}{43}{section*.71} \defcounter {refsection}{0}\relax -\contentsline {subsection}{Appendix B - Log book}{55}{section*.70} +\contentsline {subsection}{Appendix B - Log book}{56}{section*.72} diff --git a/images/interface_design.png b/images/interface_design.png new file mode 100644 index 0000000..cd92c14 Binary files /dev/null and b/images/interface_design.png differ diff --git a/report.bib b/report.bib index d08a04a..2a943d6 100644 --- a/report.bib +++ b/report.bib @@ -326,4 +326,16 @@ pages={}, year={2009}, url={http://docs.tweepy.org/en/v3.5.0/} +} + +@inproceedings{34, + title={Tensorflow Vs. Theano: What Do Researchers Prefer As An Artificial Intelligence Framework}, + author={Srishti Deoras}, + booktitle={}, + volume={}, + number={}, + pages={}, + year={2017}, + organization={Analytics India}, + url={https://www.analyticsindiamag.com/tensorflow-vs-theano-researchers-prefer-artificial-intelligence-framework} } \ No newline at end of file