Preparing Kohonen Maps for Trading Analysis with Normalization and Validation

Summary

This article updates Kohonen self-organizing map tools for algorithmic trading analysis. It replaces neuron identifiers tied to display pixels with grid indices, and adds feature normalization so inputs measured on very different scales can contribute to the learned map. A random three-feature example contrasts maps trained with and without normalization: without it, the largest-scale feature dominates; with it, structure appears across all three feature planes.

The article also describes denormalizing outputs, initializing weights around zero, setting training duration in epochs, and using a held-out validation set to monitor normalized mean squared quantization error and stop training when it begins to rise. It offers a heuristic for choosing square map dimensions from the training sample count and discusses fixes to the hexagonal grid implementation. The examples demonstrate software behavior rather than forecasting or trading performance. The proposed map-size rule is empirical, and validation and scaling choices still depend on the data and task. This installment prepares the tools; applied trading problems are deferred to a later article.

Key ideas

• Scaling input features helps prevent large-magnitude variables from dominating Kohonen map training.
• Index neurons by their map grid coordinates so display dimensions do not affect calculations.
• Standardize features using training-set means and standard deviations, then reverse the transform when interpreting outputs.
• Set training duration in epochs so each epoch presents the training samples in randomized order.
• Use held-out data to track quantization error and stop fine-tuning when validation error rises.

Tags

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