Heuristic Search for Backtesting Strategy Parameters

Summary

The document asks how to select trading and portfolio rules used in a backtest for indicators built with machine learning. The parameters include position size at different signal strengths, allocation per trade, selling rules, and whether to use a stop loss. It also raises the choice of optimization objective, such as Sharpe ratio, return, Sortino ratio, or volatility.

The author suspects ordinary minimization may be unsuitable for a non-convex objective and suggests heuristic approaches, including evolutionary and Monte Carlo search. However, the document contains no answer, experimental comparison, results, or implementation details. It therefore surfaces the design problem and candidate search families but does not establish their relative strengths, how to avoid overfitting during parameter selection, or which objective is appropriate for a given strategy.

Key ideas

• Backtest parameters can include signal-dependent position sizes, trade allocations, exit rules, and stop losses.
• Parameter selection depends on the chosen performance objective, such as return or a risk-adjusted measure.
• The author questions whether ordinary minimization is suitable for a non-convex objective.
• Evolutionary and Monte Carlo search are proposed as candidate heuristic approaches.
• The document provides no empirical evidence comparing these methods or their risks.

Tags

# Machine Learning Algorithm to find backtesting parameters?


# Machine Learning Algorithm to find backtesting parameters?












I have used ML to create indicators for performing trades in financial data. Currently I use custom logic (plenty of if conditions) for backtesting. However there must be a better approach.

Backtesting requires various parameters - how much to buy when the indicator is strong (and when it isn't), how much to sell, what percentage of portfolio to allocate in a single trade, stop loss (if any) and more. I believe they can be optimized by determining what is important (sharpe ratio or return or sortino or volatility or whatever)

I don't think normal minimization will work as this is not a convex function. But heuristic search algorithms like Evolutionary algorithm, Monte Carlo algorithm should work. However i can't find anything about this in google (if anyone has tried this before and what the results were). What defect or strength can such an approach have?

Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.