Keywords
Explainable AI, Symbolic Regression

Technologies
Python, TensorFlow, Keras, gplearn, NumPy

Stock market forecasting typically relies on black-box neural networks that offer no insight into their predictions. This project investigates whether symbolic regression can discover interpretable, closed-form mathematical equations from financial data, giving stakeholders formulas they can inspect and audit.

Two complementary approaches were implemented: a Symbolic Neural Network, whose hidden units apply named mathematical functions such as sin, multiply, and square, trained with dynamic L1 regularisation and genetic programming via gplearn, which evolves expression trees through crossover and mutation.

Both methods were validated on physics benchmarks. The Symbolic Neural Network recovered the simplified projectile motion equation exactly, and genetic programming recovered the full physical form. Applied to stock data with enriched technical indicators, the Symbolic Neural Network found no valid equation across all regularisation strengths; a result consistent with the efficient market hypothesis and itself a meaningful finding.