Selecting Trading Strategies with SVD and Neural Networks

Summary

The article describes a framework for screening candidate trading strategies by estimating their historical returns. It builds market and indicator data in MetaTrader 5, assigns directional signals to moving average, RSI, and stochastic rules, and compares those signals with subsequent price changes. The resulting strategy return series are examined with singular value decomposition to identify dominant patterns and inform a white-box selection approach; the article also describes a neural-network-based black-box alternative.

The examples use EURUSD data and show that the standalone strategies initially appear unprofitable, while the machine-selected strategy is presented as an equity curve. The text reports that the black-box result underperformed the white-box approach, with limited search iterations offered as a possible reason. The article frames the methods as ways to prioritize strategies for further testing, not proof of future profitability. Its excerpt omits parts of the analysis, and the return estimates and results depend on the selected data, horizon, indicators, and backtest design.

Key ideas

• Historical indicator data can be converted into directional strategy return series for screening.
• SVD can expose shared patterns and dominant modes among candidate strategy returns.
• The article contrasts an interpretable SVD-based selection method with a neural network approach.
• The reported black-box result trails the white-box result and may reflect limited search effort.
• Estimated historical returns are a screening aid, not evidence of future profitability.

Tags

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