Testing Neural Network Learning Rates and Class Balance for Fractal Detection

Summary

The article compares learning rates for a neural network trained to detect chart fractals. In the first experiment, separate trading programs use fixed rates or reduce the rate over time; a second experiment continues training from shared weights to control for random initialization. The trials use EURUSD hourly data, with recent candlesticks as inputs and two years of history for training. The author reports similar error near 0.42 for rates at or below 0.01, while 0.1 performs poorly. Lower error does not necessarily mean useful predictions: the low-rate model misses nearly all fractals, and reducing the rate has little effect on the reported measures.

A third experiment changes the target for bars without a fractal to partially compensate for class imbalance. The author reports fewer missed fractals and more hits after this adjustment, while describing the result as preliminary. These are limited experiments on one instrument, dataset, model, and task; the article offers no evidence that the parameter choices generalize to other markets or setups.

Key ideas

• A large learning rate can cause unstable updates, while a very small one can make training slow.
• The experiments find similar reported errors for learning rates of 0.01 and 0.001, while 0.1 performs poorly on this task.
• Error alone does not capture detection usefulness, since a model with low error can still miss nearly all fractals.
• Reducing the learning rate every ten epochs has little reported effect in these experiments.
• Changing the target for bars without a fractal improves the reported detection measures, but requires further testing.

Tags

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