Training Neural Networks to Predict Fractal Turning Points

Summary

This article builds and evaluates a supervised neural-network Expert Advisor that attempts to anticipate Bill Williams fractal turning points before the confirming candle appears. It describes choosing market inputs from historical OHLC movements, volume, time components, and several technical indicators, then feeding a rolling history into a multilayer perceptron. The network is framed in two ways: a single-output regression model that encodes buy, sell, or no signal in a signed value, and a three-output classifier for those classes. The article reports training and test statistics for the two variants, including root mean square error, hit percentage, and unrecognized fractals. After 35 epochs, the regression variant has somewhat better reported results, while both approaches show similar training time and predictive performance. These figures are specific to the described experiment and do not establish a robust trading edge. The author notes the substantial computational cost and need for training, and the forecast task is limited to fractal identification rather than a fully evaluated trading strategy.

Key ideas

• The supervised-learning target is to predict whether a confirmed fractal will appear at the previous candle's location.
• Inputs combine price shape, volume, calendar and time features, and indicator readings over a historical window.
• A regression output and a three-class output are proposed for buy, sell, and no-signal predictions.
• The reported experiment finds slightly better statistics for regression after 35 training epochs.
• Prediction scores do not by themselves demonstrate profitability, and training requires time and computational resources.

Tags

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