The Accuracy-Diversity Frontier in Adaptive Volatility Forecasting

Abstract

Adaptive forecasting systems are usually evaluated by predictive loss. This paper asks a different question: what happens to collective model diversity when forecasting systems retain different amounts of historical information? Using four equity markets from 2000–2026, a common out-of-sample period beginning in 2015, four forecast horizons, and heterogeneous forecasting technologies, I construct frozen, finite-memory, and expanding-memory forecasting cohorts that differ only in their updating rule. Full-history expanding adaptation produces lower QLIKE loss and lower effective model diversity than every finite-memory specification throughout the 16-cell market–horizon robustness grid. The cells are dependent and are not interpreted as independent replications; when horizons are collapsed within market, all four market blocks retain the same sign, but four blocks provide limited standalone statistical power. An equal-sample-size exclusion test shows that the contrast survives when broad-history and matched 3-, 5-, and 10-year learners use the same number of training observations, indicating that sample size alone does not account for the frontier. Around 22 rule-based market-specific stress onsets, expanding adaptation attenuates the post-stress loss of forecast diversity relative to a frozen benchmark; the result also survives re-clustering temporally overlapping cross-market events into global episodes. A volatility-targeting mapping shows that forecast architecture can propagate to the correlation of position adjustments, while the paper deliberately stops short of claiming a causal effect on observed market volatility. The evidence implies that predictive accuracy and collective diversity are distinct design objectives for adaptive financial forecasting systems.

Author abstract from the research manuscript. Typography and line breaks have been normalized; the linked source may show a different manuscript version.