Calibration Bets on the Past: Post-Training Quantization for Financial Time-Series Forecasting

Battle Count: 208

Highly relevant for quantitative trading firms deploying neural volatility forecasting models in production. The paper directly addresses the practical question of how to compress forecasting models for low-latency, low-memory inference without destroying predictive signal. Key findings (4-bit activation quantization causes 11-62% IC loss under default calibration; percentile calibration recovers 53-94%; recurrent/multi-scale models remain sensitive) provide actionable deployment guidelines. The walk-forward protocol mirrors realistic backtesting. However, the paper focuses on model compression rather than trading strategy construction, and does not evaluate portfolio-level P&L impact.

The core methodology (simulated quantize-dequantize, percentile calibration, walk-forward evaluation) is conceptually straightforward and implementable with standard PyTorch/TensorFlow. However, the full experimental scope (7 architectures, 8 folds, 10 seeds, 560 models, multiple calibration strategies, layerwise attribution, matched recalibration diagnostics) requires significant computational resources and careful engineering. The quantization simulation itself is simple, but reproducing the full study demands substantial GPU compute and disciplined experiment tracking.

The paper provides detailed experimental setup including dataset construction (Yahoo Finance, Tiingo), feature engineering (11 features, 64-day windows), walk-forward protocol, hyperparameter search space, training details (AdamW, batch size 1024, early stopping), and quantization implementation (simulated quantize-dequantize in FP32). However, no code repository is provided, and the specific random seeds and exact calibration sampling procedure could benefit from code release. The 560-model evaluation is computationally intensive.