Machine learning prediction of stock markets
Download as PPTX, PDF15 likes8,299 views
The document discusses the integration of machine learning in predicting financial markets, detailing various approaches such as fundamental and behavioral investing. It highlights the challenges of using machine learning in finance, including model accountability and the complexity of financial data. The author provides examples of predicting stock prices and forex market movements and emphasizes the role of sentiment analysis from social media in financial decision-making.
![Technical analysis - traditionally
• Graham criteria (1949)
• Stock Selection for the Defensive Investor:
• 1. Not less than $100 million of annual sales.
• [Note: This works out to $500 million today based on the difference in CPI/Inflation from 1971]
• 2-A. Current assets should be at least twice current liabilities.
• 2-B. Long-term debt should not exceed the net current assets.
• 3. Some earnings for the common stock in each of the past 10 years.
• 4. Uninterrupted [dividend] payments for at least the past 20 years.
• 5. A minimum increase of at least one-third in per-share earnings in the past 10 years.
• 6. Current price should not be more than 15 times average earnings.
• 7. Current price should not be more than 1-1⁄2 times the book value.
• Graham number = 22.5 ∗ 𝐸𝑃𝑆 ∗ 𝐵𝑉](https://image.slidesharecdn.com/londonml1-161119231809/85/Machine-learning-prediction-of-stock-markets-12-320.jpg)
Machine learning prediction of stock markets
- 1. Machine learning aided prediction of financial markets Nikola Milošević Email: nikola.milosevic@manchester.ac.uk Blog: http://inspiratron.org Twitter: @dreadknight011
- 2. Who am I? • BSc and MSc in EE and CS, University of Belgrade, Serbia • PhD student University of Manchester • OWASP (Open Web Application Security Project) local and project leader • Mentor at Google Summer of Code • Open Source contributor • About 4 years experience in industry (Software testing and development – Banking, Telco, SaaS, Mobile) • Head of Technology at MUTIS Finance Society
- 3. Agenda • Machine learning introduction • Finance and its resistance to machine learning • Approaches to investing • Fundamental • Behavioral • Research examples • Predicting long term stock prices using machine learning • Predicting FOREX using social media • Conclusion
- 4. What is Machine learning
- 5. How does it work? • Machine learning is the subfield of computer science that "gives computers the ability to learn without being explicitly programmed" (Arthur Samuel, 1959)
- 6. Types of Machine learning
- 8. What about handling money? • Not that much • Fund manager often can’t understand the model • Who is accountable for losses? • Data scientist/Quant/Software engineer? • Fund manager? • Data? • Algorithm? • Computer?
- 9. Machine learning as a black box or black magic
- 10. Understanding models • Machine learning ≠ black box • Machine learning ≠ black magic • Complexity matches the problem • Rules are not simple! • Complex problems – complex solutions (models) • Shortcut understanding does not work • Right cross validation, right dataset • Machines make as many mistakes as human do (in many cases even less)
- 11. Approaches in finance • H: Machine can learn same things human can Behavioural finance Long-term investing Short-term investing
- 12. Technical analysis - traditionally • Graham criteria (1949) • Stock Selection for the Defensive Investor: • 1. Not less than $100 million of annual sales. • [Note: This works out to $500 million today based on the difference in CPI/Inflation from 1971] • 2-A. Current assets should be at least twice current liabilities. • 2-B. Long-term debt should not exceed the net current assets. • 3. Some earnings for the common stock in each of the past 10 years. • 4. Uninterrupted [dividend] payments for at least the past 20 years. • 5. A minimum increase of at least one-third in per-share earnings in the past 10 years. • 6. Current price should not be more than 15 times average earnings. • 7. Current price should not be more than 1-1⁄2 times the book value. • Graham number = 22.5 ∗ 𝐸𝑃𝑆 ∗ 𝐵𝑉
- 13. Technical analysis • Analyse technical indicators and ratios • Over time • Dependencies • Graham model too strict • It is hard to develop new models • Mathematically demanding • People are not good with numbers • Time consuming • Can machines help?
- 14. Predicting long term movement of stock price • Use machine learning on past 2-3 year data • Data obtained using Bloomberg terminal • Data include 28 indicators • Book value, Market capitalization, Change of stock Net price over the one month period, Percentage change of Net price over the one month period, Dividend yield, Earnings per share, Earnings per share growth, Sales revenue turnover, Net revenue, Net revenue growth, Sales growth, Price to earnings ratio, Price to earnings ratio -five years average, Price to book ratio, Price to sales ratio, Dividend per share, Current ratio, Quick ratio, Total debt to equity, margins, asset turnover…
- 15. Predicting long term movement of stock price (2) • Selected 1739 stocks from different indexes (S&P 1000, FTSE 100 and S&P Europe 350…) • Calculated which ones price grew more than 10% • Used different Machine learning algorithms and 10 fold cross validation for evaluation • Used Python for scripting and Weka toolkit for machine learning
- 16. Results • Trial with all financial indicators as a features
- 17. Results (2) • We performed feature selection • Experiment with only 11 indicators
- 18. 11 indicators • The performance turned out not to be significantly different, but it showed that only 11 indicators are enough
- 19. FinAnalyzer • Tool that uses a model and Yahoo hidden API • https://github.com/nikolamilosevic86/FinAnalyzer • Open source (GPLv3) • Decision support • Not liable of any loses • Join development (contact me)
- 20. Behavioural finance • Psychology and emotion influence decision • People follow their peers • Sentiment of the information is shaping the decision
- 21. Social media and news
- 24. System overview
- 26. Twitter sentiment - challenges • Short text – dense information • Ungrammatical language • Abbreviations • Typos • Emoticons • Links, hashtags, mentions
- 28. Correlations – sentiment vs pair
- 29. Correlations with sentiment (What we found?) • There is correlation between sentiment and market movement • In time it is variable (some times 5 minutes, some times 5 hours) • Some times ambiguous • More research necessary • When to enter/exit?
- 30. Conclusion • Finance is all about information • Information flooded world • Machine learning, data science, text mining • Here to HELP! Try them • Predicting without emotional influence • Accountability • Consider all information for better prediction • More than human can digest
- 31. Reference • Milosevic, Nikola. “Equity forecast: Predicting long term stock price movement using machine learning. ” arXiv preprint arXiv:1603.00751 (2016). • https://arxiv.org/ftp/arxiv/papers/1603/1603.00751.pdf • Martic, Miljan. “Twitter sentiment analysis for foreign exchange market movement orediction.” (2014).