evolving trading rules with xcsf

20
Evolving Trading Rules with XCSF Daniele Loiacono Artificial Intelligence and Robotics Laboratory, Politecnico di Milano, Italy GECCO'07, July 7-11, 2007, London, UK.

Upload: daniele-loiacono

Post on 22-Jan-2015

3.367 views

Category:

Business


0 download

DESCRIPTION

 

TRANSCRIPT

  • 1. Evolving Trading Ruleswith XCSF Daniele Loiacono Artificial Intelligence and Robotics Laboratory,Politecnico di Milano, Italy GECCO'07, July 7-11, 2007, London, UK.

2. Outline

  • Introduction to XCSF
  • The framework
  • Experimental results
  • Conclusions

3. What is a learning classifier system?

  • M achine learning paradigm
  • Solves problemsevolvinga set ofrules(the classifiers)
  • Theknowledgeis thus represented with apopulationof rules
  • In stateS,doing actionAwill resultin a payoffP , estimated with accuracyF

4. XCSF

  • Introduced by Wilson in 2001
  • The classifier payoff P iscomputed by a parameterized function p(S, w )
  • Overall XCSF consists of
    • Reinforcement Learning techniques to build the payoff target (prediction)P = p(S,w)
    • Gradient descent to learn the parametersw
    • A genetic algorithm, based on accuracy, as generalization mechanism

5. Outline

  • Introduction to XCSF
  • The framework
  • Experimental results
  • Conclusions

6. The Framework STOCK10 VolumePrice STOCK1 VolumePrice STOCK2 VolumePrice INPUT ? SYSTEM Buy/Sell transactions for earning a lot of $$$ 7. Problem Decomposition STOCK10 VolumePrice STOCK1 VolumePrice STOCK2 VolumePrice INPUT Buy/Sell transactions for earning a lot of $$$STOCK1 Expert STOCK2 Expert STOCK3 Expert Merge Expert 8. The stock experts INPUT BUY HOLD SELL Action Suggested FEATURE EXTRACTION AND NORMALIZATION EXPERT KB EVOLVED WITH XCSF 9. Feature Extraction

  • The known indicators are an obvious choice
  • We used three well know indicators:
    • PPO (12,26,9) for the stock price
    • RSI with period 14 for the stock price
    • PPO (12,26,9)for the stock volume
  • In addition we also used as feature the fraction of the overall portfolio value invested in the stock
  • Finally features are normalized between -1 and 1

10. Evolving the KB

  • The KB is a set of rules evolved with XCSF applied to a single stock portfolio optimization problem.
  • At each time step XCSF can:
    • buy a fixed amount of shares of the stock
    • sell a fixed amount of shares of the stock
    • hold the current position
  • Goal is maximizing the income over time

11. Merging Expert

  • We merged all the stock expert suggestions with a simple heuristic algorithm
  • First an exponential moving average of each stock suggestions is computed (a numerical value is associated to buy, hold and sell)
  • For each stock with an average suggestion of selling higher than a threshold, a fixed amount of shares is sold
  • For each stock with an average suggestion of buying higher than a threshold, a fixed amount of shares is bought

12. Outline

  • Introduction to XCSF
  • The framework
  • Experimental results
  • Conclusions

13. Experimental setup

  • The data set provided for the competition was partitioned in two set: a training set (900 samples) and a validation set (246 samples)
  • XCSF was applied on the training set for evolving the KB of the 10 stock experts
  • A 10-fold crossvalidation was used for tuning XCSF parameters
  • Finally the overall system (the 10 stock experts and the merge one) was tested on the validation set

14. Validation results 15. Test Results

  • The system was finally tested on a the official test set provided for the competition

16. Outline

  • Introduction to XCSF
  • The framework
  • Experimental results
  • Conclusions

17. Discussion of results

  • Why so a different performance on different dataset ?
  • Financial market are complex
  • But have a look to the dataset:

TRAININGSET 18. Discussion of results

  • Why so a different performance on different dataset ?
  • Financial market are complex
  • But have a look to the dataset:

VALIDATION SET 19. Discussion of results

  • Why so a different performance on different dataset ?
  • Financial market are complex
  • But have a look to the dataset:

TESTSET 20. Conclusions

  • We introduced a very simple framework for evolving trading rules with XCSF
  • Results are mixed and further investigation is necessary for a better understanding and improvements of the system performance
  • There are a lot of possible extensions:
    • Choice of technical indicators
    • Evolving the merge expert