Automated Selection and Robustness for Systematic Trading Strategies by Dr. Thomas Starke, Quantitative Trader at Vivienne Court
Download as PPTX, PDF2 likes2,083 views
The document discusses advancements in the automated selection and robustness of systematic trading strategies, highlighting traditional approaches like parameter optimization and the challenges of overfitting. It introduces methodologies such as system parameter permutation, step-down selection, and the use of random subsamples to enhance strategy reliability while managing risks related to data mining bias. The paper emphasizes the importance of evaluating multiple parameter sets to improve performance and robustness in trading systems.
Automated Selection and Robustness for Systematic Trading Strategies by Dr. Thomas Starke, Quantitative Trader at Vivienne Court
- 1. Automated Selection and Robustness For Systematic Trading Strategies Dr Tom Starke
- 2. • Traditional approach Parameter optimisation on train set Validation on test set This is what we can do now: System Parameter Permutation Supervised compound metric optimisation Domain Regression Step-Down Correlation Minimisation Strategy Selection Automation Walk-forward optimisation
- 5. Bad result - bad luck Is there another way?
- 6. Too many parameters is bad news for systematic trading strategies But what is too many?
- 7. • Data mining bias very real Overfitting is hard to detect Solution: Use all available backtest data
- 8. System Parameter Permutation Evaluation all possible outcomes* ning Data Mining from Bias to Benefit Through System Parameter Permu
- 9. Mean should be positive PnL Distribution of all backtest results Pitfall: parameter space has unreasonable bounds
- 10. Domain Regression: The train performance informs the test performance.
- 11. Train performance informs test performance Correlation of results from backtest of train and test sets
- 12. BAD!
- 13. • Picking a single parameter set may not produce very good results Selecting several of the best strategies increases the probability of arriving closer to the regression line
- 15. Parameter sets should be as uncorrelated as possible
- 16. One More: Use random subsamples instead of fixed test period
- 17. Train Test
- 20. BRAC*- Build,Rebuild and Compare Out-of-sample performance scales with the mean of the PnL distribution of N subsamples. *Building Reliable Trading Systems - Keith Fitschen
- 21. Choosing subsamples for BRAC test
- 22. Mean PnL shortfall scales with OOS performance Performance of full data set Adjusted mean PnL of subsamples
- 25. Optimisation: Monte-Carlo vs Grid Search GA’s may not give the full story.
- 27. System Parameter Permutation Profitability Domain Regression OOS Expectation Step-Down Selection Optimising Outcome BRAC PnL Shortfall Walk-Forward Optimisation Algorithm Human-Supervised Metric Construction Output Metric Bootstrapping PnL Shortfall
- 28. Questions?