financial data mining talk

17
Can We Upgrade a Trading System with a Neural Net? Dr Michael Bowles http://www.linkedin.com/in/mikebowles [email protected]

Upload: mike-bowles

Post on 06-Dec-2014

5.276 views

Category:

Economy & Finance


4 download

DESCRIPTION

I presented these slides at a meeting of ACM data mining group. I discuss using data mining to improve performance of an existing trading system. The presentation was video taped. You can see the video at:http://fora.tv/2009/05/13/Michael_Bowles_Neural_Nets_and_Rule-Based_Trading_Systemsif you have any questions or comments contact me: [email protected] orhttp://www.linkedin.com/in/mikebowles

TRANSCRIPT

Page 1: Financial Data Mining Talk

Can We Upgrade a Trading System with a Neural Net?

Dr Michael Bowleshttp://www.linkedin.com/in/mikebowles

[email protected]

Page 2: Financial Data Mining Talk

Program

Produce a Simple Trading System Try 2 neural nets for culling trades One works the other doesn't Test whether model selection techniques

would have chosen correctly.

Page 3: Financial Data Mining Talk

1230

1235

1240

1245

1250

1255

1260

1265

1270

1275

1280

1285

1290

1230123512401245125012551260126512701275128012851290

6 :0 0 6 :3 0 7 :0 0 7 :3 0 8 :0 0 8 :3 0

E-Mini Nasdaq 100 (Continuous) (NQ #F)

2 0 0 9

Clos e Las t T op P ric e(High,10) Las t B ottom P ric e(Low,11)

T rading S trategy #2

Simple Trading System

Buy when price crosses above last high Sell when price crosses below last low

*Charts are Drawn with NeuroShell Trading Software Package

Page 4: Financial Data Mining Talk

Simple System Performance

1120

1140

1160

1180

1200

1220

1240

1260

1280

1300

0

100

200

300

400

500

D e c 2 8 J a n 4 J a n 1 1 J a n 1 8 J a n 2 5 F e b 1 F e b 8 F e b 1 5

E-Mini Nasdaq 100 (Continuous) (NQ #F)

2 0 0 8

T rading S tra tegy #2

S y s tem E quity (T rading S tra tegy #2)

Optimize over 6 weeks. Trade for 2 weeks

Page 5: Financial Data Mining Talk

Q1 Performance

Period # Return1 7.10%2 -5.20%3 6.70%4 -3.10%5 9.90%6 -3.00%7 5.10%

*In this table a “Period” is 2 weeks

Per 2 WeeksAvg Rtn: 2.50%

Rtn s.d.: 6.07%

YearlySharpe's: 2.06

Page 6: Financial Data Mining Talk

Can Statistical Learning Discard Some Losing Trades?

Give a NN some traditional indicators Train it to reject losing trades 1st try:

2 input (traditional momentum indicators) 1 to 3 hidden neurons Compare >0 or <0 for accept reject

Page 7: Financial Data Mining Talk

How does it work?

1170

1175

1180

1185

1190

1195

1200

1205

1170

1175

1180

1185

1190

1195

1200

1205

4 5 6 7 8 9 1 0 1 1 1 2 1 3 1 4 1 6 1 7 1 9 2 1 F e b 1 9 1

E-Mini Nasdaq 100 (Continuous) (NQ #F)

2 0 0 9

T rading S trategy #2

T rading S trategy

Page 8: Financial Data Mining Talk

Q1 Performance (Basically the Same)

Period # Return1 4.90%2 -3.40%3 4.00%4 -5.10%5 5.50%6 4.10%7 2.40%

Per 2 WeeksAvg Rtn: 1.77%

Rtn s.d.: 4.25%

YearlySharpe's: 2.08

*Period = 2 weeks

Page 9: Financial Data Mining Talk

Try Something Else

4 inputs (traditional momentum indicators) 2 hidden neurons Selectable tanh or Gaussian activation

functions for hiddens GA for weights and activation function

selection

Page 10: Financial Data Mining Talk

Q1 Performance (Good News / Bad News)

Period # Return1 7.80%2 -4.30%3 5.90%4 -0.50%5 7.10%6 7.30%7 5.30%

Per 2 WeeksAvg Rtn: 4.09%

Rtn s.d.: 4.64%

YearlySharpe's: 4.40

*Period = 2 weeks

Page 11: Financial Data Mining Talk

Predict the Better Performing Model?

How to judge models BEFORE looking at the OOS data

Have a look at linear regression

Page 12: Financial Data Mining Talk

2-Input Model Results(Linear Terms)

Estimate Std.Error t value Pr(>|t|)(Intercept) -0.78735 1.54228 -0.511 0.610x1 -0.14016 1.86259 -0.075 0.940x2 0.03526 0.10061 0.350 0.726

Multiple R-squared: 0.0004033AIC = 2910.813

Page 13: Financial Data Mining Talk

2-Input Model Results(Polynomial Terms)

Estimate Std. Error t value Pr(>|t|)(Intercept) -0.787343 1.549424 -0.508 0.612x1 -0.160997 2.283883 -0.070 0.944x2 0.035802 0.103914 0.345 0.731x3 0.132438 0.850780 0.156 0.876x4 0.042197 0.157572 0.268 0.789x5 0.001034 0.004287 0.241 0.810

Multiple R-squared: 0.001049AIC = 2916.614

Page 14: Financial Data Mining Talk

4-Input Model Results(Linear Terms)

Estimate StdError t value Pr(>|t|) (Intercept) -0.7929 1.5289 -0.519 0.60438 x1BB.MACD -3.1617 2.4588 -1.286 0.19948 x1BB.Price 0.7592 0.7233 1.050 0.29472 x1TT.MACD 6.8499 2.5562 2.680 0.00777 **x1TT.Price -0.9122 0.5645 -1.616 0.10712 Multiple R-squared: 0.02418AIC = 2907.399

Page 15: Financial Data Mining Talk

4-Input Model Results(Polynomial Terms)

Estimate StdError t value Pr(>|t|) (Intercept) -0.7896 1.5283 -0.517 0.60581 x1BB.MACD -2.8592 2.8054 -1.019 0.30896 x1BB.Price 0.4084 0.7980 0.512 0.60920 x1TT.MACD 6.8554 2.8108 2.439 0.01532 * x1TT.Price -1.0843 0.7179 -1.511 0.13199 x2X1x1 -3.3184 2.1810 -1.521 0.12921 x2X1x2 1.5070 1.1713 1.287 0.19927 x2X1x3 11.1495 4.1630 2.678 0.00782 **x2X1x4 -2.0146 0.8542 -2.358 0.01901 * x2X2x2 -0.1230 0.1511 -0.814 0.41623 x2X2x3 -2.6644 1.1330 -2.352 0.01935 * x2X2x4 0.4220 0.2220 1.901 0.05825 . x2X3x3 -7.5425 2.6674 -2.828 0.00501 **x2X3x4 2.9110 1.0361 2.810 0.00529 **x2X4x4 -0.2644 0.1136 -2.327 0.02066 * Multiple R-squared: 0.05708AIC = 2916.835

Page 16: Financial Data Mining Talk

Conclusion

Demonstrated NN Successfully Upgrades Trade System Statistics

Demonstrated Criteria for Model / Input Selection

Page 17: Financial Data Mining Talk

Some References

Dunis C., Laws J., Naim P, Applied Quantitative methods for trading and investment John Wiley & Sons, 2003 Ch 4. Forecasting and Trading Currency Volatility:An Application of Recurrent

Neural RegressionCh5. Implementing Neural Networks, Classification Trees, and Rule Induction

Classification Techniques: An Application to Credit Risk

Franses P.H., van Dijk D, Non-linear time series models in empirical finance, CambridgeUniversity Press, 2000Ch 5. Artificial neural networks for returns