Using an MLP and External Optimization to Generate Trading Signals

Summary

The article presents a multilayer perceptron as a transformation from market data to trading signals, separating the network's forward calculation from the algorithm that adjusts its weights. It discusses limitations of supervised, unsupervised, and reinforcement-learning approaches for market time series, then proposes evaluating candidate network weights by an aggregate trading-performance objective. Inputs can include normalized OHLC data and indicators; an external population-based optimizer searches for weights that improve the objective rather than training against labeled signal examples. The implementation is demonstrated in an Expert Advisor that processes historical data, produces directional signals, and evaluates trades using a custom profit-and-loss score. The article includes a balance curve on data described as out of sample, but the excerpt provides no detailed validation protocol or broader comparison to benchmarks. The MLP is presented as an information transformer, not as a complete learning system on its own. The EA is explicitly described as informational, with missing checks for real trade execution, so the reported illustration should not be treated as evidence of live profitability.

Key ideas

• An MLP can map market features to signals through successive nonlinear transformations.
• The proposed method selects network weights using an external optimizer and a trading-results objective instead of labeled signal targets.
• Normalized price inputs and technical indicators are used in the example EA.
• A reported out-of-sample balance curve illustrates the approach but does not establish general performance.
• The sample EA lacks checks needed for actual trading and is presented for information only.

Tags

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