Proper-score observation-driven filters: local geometry, estimation, and continuous-time limits
Giulia Livieri, Gianluca Palmari
Abstract
Observation-driven filters update a time-varying parameter with the likelihood score, linking the recursion to the logarithmic scoring rule. We replace this update with the negative parameter derivative of a differentiable proper scoring rule, within a declared working family and predictable scaling. For a general rule, the conditional mean update is a pre-conditioned stochastic-gradient of conditional scoring risk; when an autoregressive pull is included, the centre is the zero of a composite mean field. We derive local realised-loss descent and conditional-mean contraction results, and decompose the local dynamics into risk curvature and innovation variability. These two quantities coincide for the log score under the Bartlett identity but generally differ, which clarifies how bounded drivers can limit the transmission of extreme observations to the filtered path. We also establish consistency and dependent-data sandwich asymptotic normality for batch minimum-scoring-risk estimation of the static recursion parameters. For high-frequency scale models, centred updates yield a diffusion limit, non-centred updates yield a mean-flow limit, and a local transfer theorem gives an Ornstein-Uhlenbeck approximation around a moving scoring-rule risk projection. The working family, scoring rule, scaling, and autoregressive parameters jointly determine the filtered path, while the working family also determines predictive quantiles. Controlled experiments illustrate these channels. An empirical density-by-criterion factorial on international equity returns evaluates point-variance loss, value-at-risk coverage, and probability-integral-transform diagnostics without imposing a universal ranking. The results provide a criterion-based framework for observation-driven filtering under misspecification and identify the assumptions required for estimation, local tracking, and continuous-time approximation.
Read the AI summary and key takeaways for traders on WOBR Quant Research.