Filtering Moving Average Signals with Supervised Reward Models

Summary

The article develops a method for estimating whether signals from a moving average crossover are likely to be profitable. It builds on a more responsive crossover using separate moving averages on opening and closing prices, then compares prediction of the strategy’s future reward with prediction of the market’s future price change. Historical EURUSD data is described through price and indicator changes, while the original prices and indicator readings are used to label outcomes. A statistical classifier estimates whether to take or skip each signal.

The discussion uses plots to show that profitable and losing signals can occur under similar observed conditions, and describes time-series cross-validation and a comparison of learned models. It presents the approach as supervised learning applied to estimating consequences of actions, rather than as a complete reinforcement learning system. The reported results include instability in one period, which points to regime sensitivity. The method is illustrated on EURUSD and a particular crossover setup; results do not establish general profitability or transfer to other markets and strategies.

Key ideas

• A classifier can estimate whether a moving average signal is likely to produce a positive future reward.
• The strategy’s reward is defined by combining its buy or sell action with the subsequent price change.
• The article compares learning signal profitability with directly forecasting market direction.
• Overlapping profitable and unprofitable observations suggest that visual separation by indicator changes is difficult.
• Time-series validation is used, and unstable performance in a historical period highlights regime sensitivity.

Tags

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