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

Financial forecasting models are typically developed in full precision, yet production deployment often requires low-precision inference to reduce memory and computational cost. Post-training quantization (PTQ) enables such deployment without retraining, but reliable activation quantization requires calibration: activation ranges are estimated from historical data before deployment and then remain fixed during future inference. This study systematically examines activation calibration for PTQ in cross-sectional volatility forecasting on the S&P 500, covering seven representative neural architectures, eight walk-forward test years (2018-2025), and 560 trained models. The results show that activation calibration has little effect at 8 bits but becomes the primary determinant of predictive performance at 4 bits. Under default absolute-maximum calibration, static 4-bit quantization of both weights and activations removes 11-62% of the full-precision mean information coefficient in affected architectures, while percentile calibration recovers 53-94% of this degradation. The preferred activation range also varies across market periods: narrow ranges improve resolution under typical conditions but lose part of their advantage when test-period market dispersion exceeds the calibration history. These findings establish activation calibration as a first-class deployment decision for reliable 4-bit PTQ in financial forecasting, with 8-bit activations or weight-only quantization as more robust fallback choices when substantial degradation remains.