Hybrid Neural-Classical Correction for Frozen Time Series Foundation Models: A Comprehensive Ablation Study on High-Frequency Stock Prediction

Battle Count: 146

Highly relevant to quantitative trading. Directly addresses high-frequency stock return prediction during the volatile opening trading hour. Provides practical guidance on adapting foundation models for financial prediction. Key findings (classical RF matching neural correction, parameter efficiency of GatedLinear, importance of multiple correlation metrics) are directly actionable for quant practitioners. The 6.4x improvement in per-day correlation over frozen TimesFM demonstrates meaningful predictive signal. However, practical deployment would require addressing transaction costs, market impact, and broader market regime testing.

Moderate complexity. The frozen TimesFM backbone requires access to the 200M parameter model. Neural correction modules are relatively lightweight (49K-471K parameters). Random Forest is straightforward to implement. The main complexity lies in: (1) setting up TimesFM inference pipeline, (2) constructing premarket feature engineering (300x7 tensor, 21 multiscale features), (3) training the correction modules with custom loss, (4) fitting RF on residuals. The GatedLinear variant is notably simpler than AttnCorrect. No GPU-intensive fine-tuning required since base model is frozen.

Code and results publicly available on GitHub. Fixed random seed of 42 across PyTorch, NumPy, and Python's random module. Detailed hyperparameters provided (learning rates, weight decay, early stopping patience, gradient clipping, dropout). Strict chronological data splits with no shuffling. Training-only normalization. 12 model variants with clear ablation design. Open Science Statement included.