Building a B-Spline Kolmogorov-Arnold Network for Trading

Summary

The article builds a Kolmogorov-Arnold Network (KAN) that represents each connection as a learnable univariate curve. It uses cubic B-splines as basis functions, with curve coefficients fitted through ridge-regularized least squares. The network predicts the next-bar return from market features, and its learned curves are plotted to make each feature’s modeled effect visible.

The implementation includes feature construction, model storage, an indicator, and an Expert Advisor that trades predictions using a threshold and ATR-scaled stops. It describes basis and network verification and reports a small profit with shallow drawdown on an out-of-sample EURUSD M15 test over five months. The author frames this as an initial working example, not a reliable profit engine. Results depend on the selected features, fitting setup, and test window; interpretable curves show learned relationships but do not establish that those relationships will persist.

Key ideas

• A KAN learns functions on connections instead of scalar weights and fixed node activations.
• Cubic B-spline basis functions let each connection represent a flexible curve with adjustable coefficients.
• Because the model is linear in its spline coefficients, it can be fit with a direct least-squares solve.
• The example predicts next-bar returns from four market features and visualizes the learned feature curves.
• The reported out-of-sample trading result is limited evidence from one EURUSD test window.

Tags

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