Tvorba jezika japonskih svečnikov in uporaba NLP algoritma Word2Vec za napovedovanje trendov gibanja vrednosti delnic

Savić, Boris (author), Lavbič, Dejan (mentor)

V magistrskem delu rešujemo problem napovedovanja prihodnjih trendov vrednostnih papirjev s pomočjo strojnega učenja. Predstavimo nov inovativen model napovedovanja, ki temelji na uporabi japonskih svečnikov ter na NLP algoritmu Word2Vec. V delu pokažemo, da je možno iz zaporedja japonskih svečnikov tvoriti preprost jezik, katerega slovar je v primerjavi z naravnim jezikom sicer precej omejen. Z obdelavo Word2Vec sistem naučimo kontekst posameznih besed ter ta kontekst uporabimo pri gradnji napovedi. Napovedni model testiramo s simulatorjem trgovanja, ki v svojem delovanju upošteva tudi stroške trgovalne provizije. Rezultate predlaganega napovednega modela primerjamo s tremi osnovnimi modeli: Kupi in zadrži, Tekoča povprečja ter MACD. V analizi pokažemo, da predlagani napovedni model v okviru zastavljene trgovalne strategije znotraj testnega obdobja ustvari dobiček ter deluje mnogo bolje od prej omenjenih napovednih modelov. Delovanje modela preverimo tudi v validacijskem obdobju, kjer ravno tako dosežemo zadovoljive rezultate.

• Language: Slovenian
• Keywords: borza, NLP, Word2Vec, japonski svečniki, strojno učenje
• Work type: Master's thesis/paper
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2016
• PID: 20.500.12556/RUL-87581
• Publication date in RUL: 02.12.2016

Secondary language

• Language: English
• Title: The formation of Japanese candlesticks language and using NLP algorithm Word2Vec for shares trend forecasting

Abstract

Our primary objective is predicting the future trends for stock market data with machine learning. We present a new innovative prediction model, based on centuries old Japanese candlesticks and modern NLP algorithm Word2Vec. Suggested model constructs a simple language of Japanese candlesticks with a very limited vocabulary. In the following steps the prediction model uses Word2Vec to discover semantic context of each word within the language vocabulary. To test the model we develop a simulation tool that takes into account the most important aspect of stock market trading -- trade fees. To compare the success of the suggested prediction model we also develop simple TA models such as: Buy and Hold, Simple Moving Averages and MACD. Analysis of the results show the superiority of suggested prediction model over previously mentioned models in the test data-set. Additional testing is done with validation data-set in order to verify the results.

• Keywords: stock market, NLP, Word2Vec, candlesticks, machine learning