Machine learning & Time Series Analysis
•Transferir como PPTX, PDF•
1 gostou•1,045 visualizações
2019 時間序列與量化交易研討會 韓承佑博士
Recomendados
Recomendados
Mais conteúdo relacionado
Mais procurados
Mais procurados (20)
Semelhante a Machine learning & Time Series Analysis
Semelhante a Machine learning & Time Series Analysis (20)
Último
Último (20)
Machine learning & Time Series Analysis
- 1. Solutions Machine learning Time Series Analysis Challenges Finlab CTO 韓承佑 2019 / 07 / 05
- 2. Outline • Introduction • Background • Motivation • Proposed Method • Conclusion
- 3. Human Evolution Agricultural revolution Industrial revolution i Information revolution 14000 years
- 4. Human vs Computer at Go 4300 years 200 years
- 6. Gartner Hype Cycle For emerging technologies 2018 Deep Learning
- 7. Gartner Hype Cycle For emerging technologies 2015 Machine Learning
- 8. 1948 Technical Indicators 1970 Algorithmic Trading 1980 Personal Computer 1990 High frequency trading History of Trading
- 9. • Mimics the ability to see and hear • Extract rules automatically from data • ML spots patterns in high dimensional data Machine Learning
- 10. ML algorithms in finance?
- 12. What is AI, ML, DL? AI (Artificial Intelligent) ML (Machine Learning) DL (Deep Learning) @ Dog or Cat? Categories Probability Dog 0.9 Cat 0.1
- 13. Supervised Machine Learning Color Weight Age Category 3.2 kg 2 cat 4.2 kg 5 cat 6.2 kg 4 dog features labels ML Model Training Testing Color Weight Age 3.2 kg 2 4.2 kg 5 6.2 kg 4 ML Model True Answer Prediction cat cat cat dog dog dog
- 14. Outline • Machine Learning Models • Training & Testing • Evaluation • Feature Engineering • Data Preprocessing
- 15. Feature Engineering & Data Preprocessing
- 16. Feature Source • Difficult to confirm data release date • Missing data is often backfilled • Consider multiple correction • Maybe useful to combine other data types Fundamental data • Trading book • market participant characteristic footprint • Massive amount of data generated in one day Market data Alternative data • news, trend, web, satellites… • Primary data source • Hard to process, difficult to confirm consistency
- 17. Challenging of Labeling the data Time t KD RSI MACD Category 1 -1 2 -1 … 0 t 1 features labels price time 1 0 -1 𝑝 𝑡 + 𝑤 + 𝜏 𝑝 𝑡 + 𝑤 − 𝜏 𝑝 𝑡 + 𝑤𝑝 𝑡 • A popular method in the literature • 𝜏 is a constant regardless of the volatility • Do not have stop-loss limits Fixed time horizon
- 18. Labeling generation for financial price • Triple barrier [Prado 2018] • Continuous trading signals [Dash 2016] • Trading Point decision [Chang 2009] [Prado 2018] Advances in Financial Machine Learning [Tsantekidis 2017] Using Deep Learning to Detect Price Change Indications in Financial Markets [Dash 2016] A hybrid stock trading framework integrating technical analysis with machine learning techniques [Chang 2009] Integrating a Piecewise Linear Representation Method and a Neural Network Model for Stock Trading Points Prediction
- 19. Triple barrier [Prado 2018] price time 1 0 -1 𝑝 𝑡 + 𝑤𝑝 𝑡 𝑝 𝑡 + 𝑤 + 𝜏1 𝑝 𝑡 + 𝑤 − 𝜏2 • Labels according to the first barrier touched out of three barriers • horizontal barriers are defined by profit-taking and stop- loss limit • 𝜏1 and 𝜏2 are dynamic according to estimated volatility
- 20. Continuous trading signals [Dash 2016] time price 𝑝 𝑡 + 𝑤𝑝 𝑡 𝑝𝑡,𝑡+𝑤 max 𝑝𝑡,𝑡+𝑤 min 𝑝𝑡,𝑡+𝑤 min 𝑝𝑡,𝑡+𝑤 max 0.5 1 0.5 0 • Using momentum of the stock price • y(t)’s are continuous • Provides more detailed information
- 21. Trading point decision [Chang 2009] • Find the local minimum and maximum points • Divide the time series into subsegments • Threshold value d length of trend
- 22. Trading point decision [Chang 2009] • Find the local minimum and maximum points • Divide the time series into subsegments • Threshold value d length of trend
- 23. Trading point decision [Chang 2009] • Find the local minimum and maximum points • Divide the time series into subsegments • Threshold value d length of trend
- 24. Trading point decision • Find the local minimum and maximum points • Divide the time series into subsegments • Threshold value d length of trend
- 25. Trading point decision • Find the local minimum and maximum points • Divide the time series into subsegments • Threshold value d length of trend
- 27. Neural Network [McCulloch 1943] • Built to model the human brain • Designed to recognize patterns • Interpret numeric data through a kind of machine perception Human neuron structure Single neuron model y1= g (w1x1+w2x2+ v g(v) 1 w0 w1x1 x2 g w2 y1 [McCulloch 1943] A Logical Calculus of Ideas Immanent in Nervous Activity
- 28. Neural network [McCulloch 1943] Single node in neural network g y1 1 w0 w1x1 x2 w2 v g(v)
- 29. Neural network [McCulloch 1943] Simplified expression y1 1 w0 w1x1 x2 w2
- 30. Neural network [McCulloch 1943] A layer contain multiple neurons 1 x1 x2 y2 y1 y3 y4
- 31. Deep Neural network Multi-layer deep neural network 1 x1 x2 y2 y3 y1
- 32. Deep Neural Network Training Result 2018-1-1 2019-7-1 Train 2006 ~ 2014 2016 ~ 2019-3-1 Validate 2015 Backtest Features Scaled Technical Indicators Asset Data split Labels Fixed time horizon Taiwan Capitalization Weighted Stock Index benchmark backtest
- 33. Long short term memory neural network(LSTM) [Hochreiter 1997] • Can process sequence of data • LSTM deals with the exploding and vanishing gradient problems Hochreiter, Sepp, and Jürgen Schmidhuber. "Long short-term memory." Neural computation 9.8 (1997): 1735-1780. Neural network x y Neural network x y Neural network x y Neural network x y Neural network x y t = 1 t = 2 t = 3 x y Input gate + tanh LST M Forget gate Output gate
- 34. Long short term memory neural network(LSTM) [Hochreiter 1997] Train 2006 ~ 2014 2016 ~ 2019-3-1 Validate 2015 Backtest Features Scaled Technical Indicators Asset Data split Labels Fixed time horizon Taiwan Capitalization Weighted Stock Index loss profit epoch
- 35. Convolutional Neural Network [2012 Krizhevsky] • Commonly applied to computer visual imagery • Prevent overfitting Convolutional Layer
- 36. Time series to Image conversion Approach [Sezer 2018] [Sezer 2018] Algorithmic FinancialTrading with Deep Convolutional Neural Networks:Time Series to Image Conversion Approach
- 37. Which time series is random walk? 1 2 3 4 5 6
- 38. Generative Adversarial Networks (GAN) Historical Price data Real/Generated • The Generator is trained to generate data that looks like historical price • The Discriminator is trained to tell the difference between generated and real data https://github.com/nmharmon8/StockMarketGAN
- 39. StockMarketGAN Discriminator features is good for predict the direction of the price Historical Price data Fixed Weights (New features) Rise/Fall https://github.com/nmharmon8/StockMarketGAN
- 40. Backtest • Survivor bias, lookahead bias, training, transection cost, outlier • Finding the lottery tickets that won the last game • Machine learning overfitting • Solutions • Develop model for entire. asset or classes • Use Bootstrap aggregating • Record every backtest conducted • K-fold Cross validation
- 41. K-fold Cross Validation • Determine the generalization error of an ML algorithm • Prevent overfitting • Assume the training set and the testing set are IID
- 42. Drawback Train Test Train Price • Training set contains information that also appears in the testing set • Observations cannot be assumed to be drawn from an IID process • Multiple testing and selection bias
- 43. Purged K-fold Cross Validation [Prado 2018] Price Price Before Purging After Purging
- 44. Feature Importance [Liaw 2002] • Understand features contributed to the performance • Add some features that strengthen the predictive power • Opens up the proverbial black box • How to deal with selection bias • Evaluate features in multiple assets • Using different tree base classifier/regressor • Use random feature to distinguish powerful features
- 45. Mean decrease impurity (MDI) • Adding up information gain for each features • Simple and efficient to calculate • inflate importance of continuous feature Correlation Coefficient 0.94 Importance indicator ID’s Humidity Rain Sunny day Throw a coin Rain Sunny day Importance: technical indicators of TXF1 Importance: technical Indicators of random walk time series [2013 Louppe] Understanding variable importances in forests of randomized trees. [2013 Louppe]
- 46. Permutation Importance • Much more computationally expensive • Results are more reliable Color Weight Age 3.2 kg 2 4.2 kg 5 6.2 kg 4 baseline = model.score(X, y) for each column in X permutate column in X randomly reduced_score = model.score(X, y) recover the column importance of the feature = baseline – reduced_score indicator ID’s Importance: technical indicators of TXF1 Importance: technical Indicators of random walk time series -0.04 Correlation Coefficient Importance [2013 Louppe] [2013 Louppe] Understanding variable importances in forests of randomized trees
- 47. Conclusion • Feature engineering • Labeling • Fixed time horizon, triple barrier, Continuous signal • Multi-asset feature selection • MDI, MDA • Model • Neural network • LSTN • CNN • Cross validation, backtest • Realistic check
- 48. Future work Feature enumeration Stationarize features Feature selection Preprocessing Training Trading strategies development Reality check Multiple assets Strategies Strategy management Live trading
- 49. Q&A Facebook AI CourseFinlab Blog AI Diagnose
Notas do Editor
- 好幾千年的進化， 解決生存問題 - 農業 解決勞動問題 - 工業革命 解決思考問題 - 資訊革命 人類是很懶的生物 懶得出外打獵 懶得勞動 懶得思考 人類思考了這麼久
- 我們如今走到了這裡，200年前真空管發明到現在，電腦取得了巨大的進步
- 要是…就好了 要是…很痛苦 要是…躺著賺
- As emotional beings, subject to fears, hopes, and agendas, humans are not particularly good at making fact-based decisions, particularly when those decisions involve conflicts of interest. Millions of years of evolution (a genetic algorithm) have fine-tuned our ape brains to survive in a hostile 3-dimensional world where the laws of nature are static. Humans are slow learners, which puts us at a disadvantage in a fast-changing world like finance. 10 year is very different The first algorithm trading Technical Trading -> Quantitive trading -> high frequency trading -> machine learning trading