Deployment-Side Adaptiveness in Multi-Horizon Volatility Forecasting

A new study finds that how a trained financial model is deployed at inference time can significantly alter its forecasting accuracy, challenging the common practice of evaluating models solely on their architecture. The research, posted to arXiv on June 26, examines multi-horizon volatility forecasting and argues that a trained multi-output (MIMO) forecaster does not produce a single fixed predictor [1]. By changing the inference-time rollout rule, the same trained model generates a family of forecasts with different accuracy and computational cost profiles [2]. The authors tested this across 20 stock-volatility series and three forecast horizons, using architectures that ranged from linear models to the transformer-based PatchTST [2]. They found that non-default rollout rules often improved over standard MIMO deployment, but the best fixed rule varied substantially across architectures and horizons, making any single static replacement unreliable [2]. To address this, the study evaluated validation-based deployment policies over the induced rule family. Under the primary mean squared error (MSE) objective, validation-selected singletons provided a low-cost improvement over default MIMO, while small rule subsets recovered much of the benefit of larger ensembles at substantially lower inference cost [2]. The work also revealed that policy rankings are metric-sensitive. Policies selected to optimize MSE did not transfer uniformly to QLIKE, a finance-standard volatility loss function [2]. This finding underscores that deployment choices must align with the specific evaluation metric used in practice. The study’s focus on inference-time adaptiveness arrives amid a period of heightened real-world volatility. The 2026 Iran war and the subsequent closure of the Strait of Hormuz triggered what the International Energy Agency characterized as the “largest supply disruption in the history of the global oil market,” causing Brent Crude to surge past $120 per barrel and forcing QatarEnergy to declare force majeure on all exports [3]. Such shocks place extreme demands on financial forecasting systems, where model deployment decisions can directly affect risk assessments. The authors conclude that trained volatility forecasters should be evaluated not only by their architecture, but also by their deployment policy [1]. The paper was released on arXiv under the machine learning category and is associated with the arXivLabs framework for experimental community projects [1].