When Does Volatility Model Selection Matter? Entropy Diagnostics and Pre-Registered Evidence Across 1,496 Assets and Eleven Asset Classes

We study when volatility model selection has value, rather than which model wins on average. A single entropy statistic, computable in seconds from a trailing return window, identifies 94% of assets where a simple EWMA baseline suffices—eliminating 86% of model-fitting cost. A 2×2 in-sample attribution separates two mechanisms: filtered historical simulation (FHS) fixes unconditional VaR coverage, while model diversity reduces violation clustering. Walk-forward backtesting across 1,491 assets establishes the paper's central out-of-sample result: per-window best-model selection overfits, but forecast combination (EW-COMB) preserves the Christoffersen clustering benefit (+6.8–8.3 percentage points over EWMA+FHS, p < 10⁻¹¹). Model selection does not improve out-of-sample volatility forecasting or generate position-sizing alpha, bounding the framework's value to computational triage, regime diagnostics, and OOS violation-clustering reduction via forecast combination. The framework evaluates twelve volatility forecasters across a 1,496-asset, 11-class cross-asset universe with twelve pre-registered hypotheses, cryptographic spec-locking, and hash-chained computation records. Nine hypotheses pass after multiple-testing corrections; three null results are reported with equal rigor. Twenty-eight supplementary analyses (S1–S28) are released as an extensible interface against the benchmark's immutable SQLite data store, enabling researchers to test new hypotheses without modifying the core pipeline or breaking the pre-registration chain. Repository: https://github.com/oliviersaidi/PACF_F License: CC BY-NC-SA 4.0