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