Deployment Strategies Boost Multi-Horizon Volatility Forecasting Performance.

In financial forecasting, the predictive performance of a model isn't solely determined by its architecture or training. A crucial, yet often overlooked, factor is how the trained model is deployed. This research investigates "deployment-side adaptiveness" in the context of multi-horizon volatility forecasting, a critical task for financial markets. The study highlights that a single trained multi-output (MIMO) forecaster can generate a family of forecasts, each with different accuracy and cost profiles, simply by altering its inference-time rollout rule. Across 20 stock-volatility series, various forecast horizons, and diverse model architectures (from linear models to PatchTST), it was found that non-default rollout rules frequently outperform standard MIMO deployment. However, the optimal fixed rule varies considerably across different architectures and horizons, making a static replacement unreliable. Consequently, the researchers evaluated validation-based deployment policies over this family of rules. They discovered that validation-selected single rules provide a low-cost improvement over default MIMO, while small subsets of rules can capture much of the benefit of larger ensembles with significantly reduced inference costs. The study also notes that policy rankings are metric-sensitive; policies optimized for Mean Squared Error (MSE) do not necessarily transfer uniformly to other finance-standard metrics like QLIKE. These findings underscore that inference-time deployment is a vital source of adaptiveness in financial forecasting, suggesting that models should be evaluated not just by their training but also by their deployment strategy.