Transcript PPTX
Currency Forecasting using Multiple Kernel Learning with Financially Motivated Features Tristan Fletcher, Zakria Hussain and John Shawe- Taylor Fanghua Lin Financial Services Analytics Content β’ Motivation β’ Empirical Study β’ Results β’ Conclusion & Contribution Motivation A trader can profit from accurate prediction of currency trend: Three situations (buying price (bid) and selling price (ask) ): If π΅ππ ππ‘+Ξπ‘ > ππ‘π΄π π (> ππ‘π΅ππ ) π©ππ ππππππππ π΄π π π΅ππ If ππ‘π΅ππ > ππ‘+Ξπ‘ (> ππ‘+Ξπ‘ ) Sell ππππππππ π΅ππ π΄π π If ππ‘+Ξπ‘ < ππ‘π΄π π πππ ππ‘π΅ππ < ππ‘+Ξπ‘ Do nothing Financially Motivated Features Price-based features β’ β’ β’ β’ F1 = F2 = F3 = F4 = EMAπΏ1 , β¦ , EMAπΏπ MAπΏ1 , β¦ , MAπΏπ , Οπ‘ πΏ1 , β¦ , Οπ‘ πΏπ ππ‘ , maxπ‘ πΏ1 , β¦ , πππ₯π‘ πΏπ , minπ‘ πΏ1 , β¦ , minπ‘ πΏπ βπ‘ πΏ1 , β¦ , βπ‘ πΏπ , βπ‘ πΏ1 , β¦ , βπ‘ πΏπ Volume-based features β’ F5β¦8 = ππ‘ , ππ‘ ,π ππ‘ 1 π‘ β ππ‘β1 , ππ‘ βππ‘β1 ππ‘ βππ‘β1 1 β’ Experimental Design: β’ Ξ1:5 = exp(β π₯ β π₯ , 2 /Ο1 2 ), β¦ , exp(β π₯ β π₯ , β’ Ξ 6:10 = (< π₯, π₯ , > +1)π1 , β¦ , (< π₯, π₯ , > +1)π5 β’ Ξ11:15 = 2 β1 sin π β’ Ξ16 = < π, π, > 2ππ 1 π₯ , 1+2π₯ π 1π₯ 1+2π₯ ,π 1 π₯, 2 /Ο 2 ) 5 2 , β¦ , π sinβ1 ( 2π₯ π 5 π₯ , (1+2π₯ π 5 π₯)(1+2π₯ ,π ) 5 π₯,) Empirical Study 8*16=128 feature/kernel combinations Fπ Ξπ : the combination of i feature with j kernel Three SVM are trained on the data: π΅ππ SVM 1: ππ‘+Ξπ‘ > ππ‘π΄π π π¦π‘1 = +1, ππ‘βπππ€ππ π π¦π‘1 = β1 π΄π π SVM 2: ππ‘π΅ππ > ππ‘+Ξπ‘ π¦π‘2 = +1, ππ‘βπππ€ππ π π¦π‘2 = β1 π΅ππ π΄π π SVM 3: ππ‘+Ξπ‘ < ππ‘π΄π π πππ ππ‘π΅ππ < ππ‘+Ξπ‘ π¦π‘3 = +1, ππ‘βπππ€ππ π π¦π‘3 = β1 ππ = π¦π‘1 , π¦π‘2 , π¦π‘3 = ±π, ±π, ±π , ππ is correct, when only one element of ππ is postive Empirical Study Training: 100 Testing:100 Time 100 days shifting 100 days shifting Training:100 Testing:100 Time 10-fold cross-validation was used to select the three kernels with the highest predictive accuracy for the dataset, namely F8 Ξ16 , F1 Ξ1 πππ F1 Ξ 3 Results Results Percentage Accuracy of Predictions Ξt: Time Horizon (Prediction) 5 10 20 50 100 200 MKL 94.7 89.9 81.7 67.1 61.1 58.9 F8K16 94.7 89.6 81.3 65.4 51.1 45.0 F1K1 93.0 88.4 79.5 65.5 60.7 28.8 F1K3 92.8 84.6 72.3 61.1 59.9 61.3 Conclusion The most successful individual kernels are selected by cross-validation are awarded very low weights by SimpleMKL. This reflects a common feature of trading rules where individual signals can drastically change their significance in terms of performance when use in combination. Furthermore, the effective method of combining a set of price and volume based features in order to correctly forecast the direction of price movements in a manner similar to a trading rule Financial Forecasting with Gompertz Multiple Kernel Learning Han Qin Dejing Dou Yue Fang 2010 IEEE International Conference on Data Mining Fanghua Lin Financial Services Analytics Content β’ Models β’ Garch β’ Gompertz Function β’ Gomperz Multiple Kernel Learning β’ Subgradient Descent Algorithm β’ Empirical Study β’ Conclusion & Contribution Garch Model Garch Model: Οπ‘ 2 = πΌ0 + π 2 πΌ R π=1 π π‘βπ + Return of Stock π 2 π½ Ο π π‘βπ π=1 πππππ‘ππππ‘π¦ ππ ππ‘πππ i.e., Future Volatility = π(Past Returns, Past Volatilities) Kernel Function Gompertz Function Assigns higher weights to most recent data Garch Model Gompertz Function SVM Subgradient Descent Algorithm Gomperz Multiple Kernel Learning (GMKL) Difference between LMKL and GMKL ο· LMKL :training data and test data have same distributions. GMKL addresses the non-stationary problem by favoring recent data. ο· LMKL :single data source but different kernel functions. GMKL: different data sources with same kernel function. ο· LMKL : discovers which kernel function is better for a certain region of the kernel matrix. GMKL: assigns the weights to different regions by considering the order of time series data. Empirical Study Data Index Daily Index Closing Price οΌeg, General Motors CorporationοΌ 5 major international stock indexes: Time Period Goal Comparsion Dw Jones Industrial Average, S &P 500, FTSE 100, Hengsheng , Nikei 225 Jan 2007 β Dec 2009 Model 1: SVM Model 2: MKL Forecasting Accuracy Model 3: GMKL Relative Absolute Error (RAE) Test Many Shifting Periods and Average the Performance Testing Training n day shifting Time n day shifting Training Testing Time Forecasting DJI using DJI and one other index Forecasting using all 5 indexes Conclusion & Contribution Conclusion: β’ GMKL performs better than both SVM and MKL. β’ GMKL model is more robust than MKL when considering more training data sources Contribution: For data mining: β’ novel model to integrate multiple financial time series data sources β’ Propose a domain specific kernel function to leverage domain knowledge in the mining process For financial forecasting: β’ New method to tackle the international market integration problem β’ address the non-stationary of the financial time series data β’ Reveal interesting relationships among multiple international stock markets Start-up: Thought Machine Thought Machine is building technology to revolutionize the way people do their day to day banking. Using Machine Learning to analyze transactions, to find patterns and let users better understand and manage their finances. CEO: Paul Taylor β’ Working Experience: Manager and Technical Lead in Google, Chief Executive Officer in Phonetic Arts, Visiting Lecture in University of Cambridge β’ Education: PhD, Edinburgh Universityβs Centre for Speech Technology Research