41x Knowledge Distillation from Recurrent Ensembles to Liquid Continuous-Time Networks for Financial Time Series Classification

We present a knowledge distillation framework that compresses a 137,763-parameter recurrent ensemble (LSTM + GRU) into a 3,328-parameter Liquid Continuous-time Closed-form (LiquidCfC) network, achieving 41.4x parameter reduction for directional classification on forex time series. The teacher achieves a pooled bootstrap Sharpe ratio of 71.4 (95% CI: [58.4, 85.1]) while the student achieves 16.3 (95% CI: [14.76, 17.89]), demonstrating statistically significant predictive fidelity under extreme compression. We introduce a deployment gate criterion (G3) based on KL divergence at unit temperature, providing a principled accept/reject mechanism for compressed model deployment. Validated via Monte Carlo block bootstrap (10,000 resamples) on expanding-window walk-forward cross-validation across four major currency pairs.