Leakage-Controlled Chronological Evaluation of Hybrid Option Pricing Models

Abstract

Option-pricing comparisons can be biased when quote-derived inputs or later outcomes enter model selection. We compare analytical, learned, fixed-blend, and conditional-blend models in a chronology that ends tuning in 2019 and fixes the 2024 to 2025 confirmation before its outcomes are examined. SPY, QQQ, and IWM are assessed in four prespecified sampling streams using raw mean-daily normalized-midpoint MAE. Every model uses features available at prediction time and is scored on the same contract-date keys within each asset-stream cell. The MAE-aligned mixture improves on the validation-optimal fixed blend in 1/4 SPY, 4/4 QQQ, and 4/4 IWM streams; SPY also has reversals and unresolved intervals. Across all 12 asset-stream cells, a simpler winner-probability soft gate has lower macro-average error, while component removals show no stable direction. Date-cluster intervals and a moving-block sensitivity check retain that uncertainty and report unfavorable streams alongside favorable ones. Conditional blending is thus asset- and sample-dependent, not uniformly preferable. The study contributes a leakage-controlled evaluation protocol and a bounded empirical comparison.