Ensembling Neural Network Regressors by Averaging Predictions

Summary

This implementation builds a committee of neural network regressors, trains each member on the same training data with validation data and early stopping, then averages their predictions. The model class and parameters, committee size, training epochs, and stopping patience are configurable. It also retains training histories and plots each member’s training and validation loss.

Averaging several model outputs is intended to combine their estimates, but the code provides no empirical comparison against a single model or other ensemble methods. It does not describe how member diversity is created, how predictions should be converted into trading decisions, or how the approach performs out of sample. Consequently, it is an implementation pattern for model aggregation rather than evidence of a profitable trading strategy.

Key ideas

• The committee trains a configurable number of neural network regressors.
• Each model uses validation loss for early stopping.
• The returned prediction is the mean across committee members.
• Training and validation loss histories can be plotted for individual models.
• The document provides no evidence of improved trading or forecasting performance.

Tags

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