Proper-Score Observation-Driven Filters: Local Geometry, Estimation, and Continuous-Time Limits

Battle Count: 50

Highly relevant for quantitative trading applications involving volatility forecasting, VaR calibration, and robust filtering of time-varying parameters. The paper directly addresses how criterion choice affects value-at-risk calibration, variance-forecast loss, and adaptation to structural breaks—all critical for risk management in trading desks. The empirical application on twelve international equity indices demonstrates practical relevance. The robustness results (bounded vs. unbounded drivers) inform how trading systems should handle outlier returns. However, the paper is primarily theoretical/methodological rather than providing a ready-to-deploy trading strategy.

Moderate to high complexity. The theoretical framework requires understanding of proper scoring rules, stochastic approximation, and continuous-time limits. Implementation requires: (1) computing rule-specific derivatives and scalings for multiple scoring rules, (2) handling the autoregressive recursion with proper initialization, (3) implementing batch minimum-scoring-risk estimation with L-BFGS-B and multistart, (4) computing PIT diagnostics, VaR coverage, and QLIKE losses, (5) handling edge cases like infinite moments for log-score under heavy tails. The closed-form updates for Gaussian density make the core recursion tractable, but MMD requires quadrature and the full empirical protocol involves extensive diagnostics and bootstrap inference.

The paper provides detailed simulation designs with explicit seeds (20260430), replication counts, parameter grids, and complete protocols in Appendix F. All scoring rule formulas, scaling choices, and tuning constants are explicitly stated. The empirical application uses publicly available data (Yahoo Finance, FRED). However, no code repository is linked, and some numerical procedures (e.g., MMD quadrature, L-BFGS-B optimization with 8 starts) require careful implementation. The theoretical proofs are complete in appendices.