Deployment Strategies Crucial for Volatility Forecasting Performance

In financial forecasting, the predictive performance of a model isn't solely determined by its training; how it's deployed during inference also plays a critical role. This study investigates this aspect within multi-horizon volatility forecasting. A trained multi-output (MIMO) forecaster can generate various forecasts depending on the inference-time rollout rule applied, each with distinct accuracy and cost profiles. Across 20 stock-volatility series, three forecast horizons, and diverse architectures (from linear models to PatchTST), non-default rollout rules frequently improved upon standard MIMO deployment. However, the optimal fixed rule varied considerably across architectures and horizons, making a single static replacement unreliable. The researchers evaluated validation-based deployment policies over the induced rule family. For the primary Mean Squared Error (MSE) objective, validation-selected single rules offered a low-cost improvement, while small subsets of rules captured much of the benefit of larger ensembles at lower inference cost. The study also found that policy rankings are metric-sensitive, meaning MSE-selected policies did not uniformly transfer to QLIKE, a standard finance volatility loss. This highlights that inference-time deployment is a significant source of adaptiveness in financial forecasting, and trained volatility forecasters should be assessed by both their architecture and their deployment policy.