Tweedie's Formula and Score-Driven Updating
Pith reviewed 2026-05-19 18:19 UTC · model grok-4.3
The pith
Score-driven updates match Bayesian posterior corrections exactly for conjugate natural exponential families under steady-state precision discounting and inverse-Fisher scaling.
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
Core claim
For conjugate natural exponential families, the classical discounted Bayesian recursion has an exact score-driven representation: with steady-state precision discounting and expectation-space inverse-Fisher scaling, the score-driven correction equals the Bayesian posterior mean before transition dynamics are imposed. Tweedie's variance-function index further clarifies how conditional scores normalize forecast errors. The results link empirical Bayes, approximate filtering, dynamic generalized linear models, and score-driven models while distinguishing exact Bayesian updating from local score-based approximation.
What carries the argument
Tweedie's formula that relates the posterior mean correction to the score of the marginal predictive density, including a base-measure adjustment in natural exponential families; it provides the identity that makes the score-driven correction match the Bayesian update.
If this is right
- Score-driven models admit an exact Bayesian reading in conjugate natural exponential families.
- Conditional scores normalize forecast errors according to the variance function.
- Local Gaussian approximations via Fisher scoring work for general densities outside the exact cases.
- Practitioners gain a way to choose between exact Bayesian and score-driven recursions based on the model family.
Where Pith is reading between the lines
- This equivalence may allow importing computational or theoretical results from Bayesian filtering into score-driven applications.
- Future work could examine how well the local approximation performs when the density is not from a natural exponential family.
- Modelers in econometrics might prefer the score-driven form for its simplicity when the Bayesian interpretation holds.
Load-bearing premise
The derivations require conditional densities to belong to natural exponential families for exact equivalence and Fisher scoring to produce a valid local Gaussian posterior correction for general densities.
What would settle it
In a conjugate family such as the Poisson, compute both the exact discounted Bayesian posterior mean update and the score-driven correction with inverse-Fisher scaling and check if they coincide before any state transition.
read the original abstract
Score-driven models update time-varying parameters using conditional likelihood scores. This paper develops a Bayesian interpretation of such updates through Tweedie's formula, which connects posterior mean corrections with marginal scores. In Gaussian signal extraction, this gives an exact posterior-correction identity. For natural exponential families, related identities characterize posterior means in natural- and expectation-parameter spaces. Building on these identities, we show that conjugate Bayesian filtering in expectation space coincides exactly with an inverse-Fisher-scaled conditional score update under local precision discounting. For general conditional densities, the exact Bayesian correction involves a generally unavailable predictive-marginal score. A local Gaussian approximation shows that the conditional likelihood score provides the leading approximation to this posterior correction; under local precision discounting, the predictive covariance becomes proportional to inverse Fisher information, yielding the familiar inverse-Fisher-scaled score recursion. The results clarify when score-driven updates are exact Bayesian filters and when they should instead be viewed as tractable local approximations.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper claims that score-driven models for time-varying parameters admit a Bayesian interpretation via Tweedie's formula. In the Gaussian case the posterior correction equals a scaled score of the marginal predictive density; in natural exponential families the identity includes an explicit base-measure adjustment. For general conditional densities, inverse-Fisher-scaled conditional scores arise as local Gaussian posterior corrections obtained by Fisher scoring together with precision discounting. For conjugate natural exponential families the classical discounted Bayesian recursion possesses an exact score-driven representation once steady-state precision discounting and expectation-space inverse-Fisher scaling are imposed; under these choices the score-driven correction equals the Bayesian posterior mean before any transition dynamics are applied. Tweedie's variance-function index is shown to clarify normalization of forecast errors. The results are presented as linking empirical Bayes, approximate filtering, dynamic generalized linear models, and score-driven updating while distinguishing exact Bayesian recursion from local score-based approximation.
Significance. If the claimed exact identities hold, the paper supplies a theoretically grounded bridge between score-driven time-series methods and Bayesian filtering that could justify the use of score-driven recursions inside dynamic GLMs and improve interpretability of updates in econometric applications. The variance-function perspective on forecast-error normalization is a useful ancillary contribution. The work is strongest where it delivers parameter-free or exact equivalences rather than approximations.
major comments (2)
- [Section deriving the exact representation for conjugate natural exponential families] The central claim that, for conjugate NEFs, steady-state precision discounting plus expectation-space inverse-Fisher scaling produces an exact identity between the score-driven correction and the Bayesian posterior mean (before dynamics) is load-bearing. The derivation must demonstrate explicit cancellation of the Tweedie score term, the base-measure adjustment, and the chosen scaling; without an independent symbolic check or explicit expansion for at least one non-Gaussian conjugate family (e.g., Poisson or Gamma), the step converting the classical discounted Bayesian recursion into the scaled score remains the least secure link.
- [Section on general conditional densities and Fisher scoring] In the general-densities case the manuscript invokes Fisher scoring to obtain local Gaussian posterior corrections. It is unclear whether the precision-discounting step is applied before or after the local Gaussian approximation; this ordering affects whether the resulting inverse-Fisher scaling remains exact or becomes an additional approximation. Clarify the sequence and state the conditions under which the representation continues to hold.
minor comments (2)
- [Abstract] The abstract is information-dense; splitting the long sentence that begins 'For conjugate natural exponential families...' would improve readability.
- [Notation and definitions] Notation for the variance-function index and for the expectation-space Fisher information should be introduced once and used consistently; cross-references to the relevant equations would help.
Simulated Author's Rebuttal
We thank the referee for the careful reading and constructive comments. The points raised help clarify the presentation of the exact identities and the local approximations. We address each major comment below and will incorporate the suggested revisions.
read point-by-point responses
-
Referee: [Section deriving the exact representation for conjugate natural exponential families] The central claim that, for conjugate NEFs, steady-state precision discounting plus expectation-space inverse-Fisher scaling produces an exact identity between the score-driven correction and the Bayesian posterior mean (before dynamics) is load-bearing. The derivation must demonstrate explicit cancellation of the Tweedie score term, the base-measure adjustment, and the chosen scaling; without an independent symbolic check or explicit expansion for at least one non-Gaussian conjugate family (e.g., Poisson or Gamma), the step converting the classical discounted Bayesian recursion into the scaled score remains the least secure link.
Authors: We agree that an explicit verification strengthens the central claim. The manuscript derives the general identity for conjugate NEFs by showing how the chosen scaling and discounting produce cancellation between the Tweedie score, base-measure term, and posterior-mean update. To address the concern directly, the revised version will include an appendix with the full symbolic expansion for the Poisson conjugate family (and, if space permits, the Gamma case), confirming term-by-term cancellation under steady-state precision discounting and expectation-space inverse-Fisher scaling. revision: yes
-
Referee: [Section on general conditional densities and Fisher scoring] In the general-densities case the manuscript invokes Fisher scoring to obtain local Gaussian posterior corrections. It is unclear whether the precision-discounting step is applied before or after the local Gaussian approximation; this ordering affects whether the resulting inverse-Fisher scaling remains exact or becomes an additional approximation. Clarify the sequence and state the conditions under which the representation continues to hold.
Authors: We thank the referee for noting the potential ambiguity in ordering. In the framework, precision discounting is applied to the prior precision before the local Gaussian approximation is obtained via Fisher scoring; the inverse-Fisher scaling is then taken with respect to the already-discounted precision. This sequence keeps the scaling exact within the local approximation. The revised manuscript will state this ordering explicitly in the relevant section and add a brief remark on the conditions (local validity of the quadratic approximation around the current parameter value) under which the representation holds. revision: yes
Circularity Check
Derivation from Tweedie's formula and Fisher information is self-contained with no reduction to inputs by construction
full rationale
The paper starts from the established Tweedie's formula for posterior corrections in Gaussian and natural exponential family settings, then shows that inverse-Fisher-scaled scores arise as local Gaussian approximations via Fisher scoring and precision discounting. For conjugate NEFs the exact equivalence to the discounted Bayesian posterior mean (before dynamics) follows from algebraic cancellation under the stated scalings. No step renames a fitted quantity as a prediction, imports uniqueness from the authors' prior work, or defines the target result in terms of itself. The central identity is a derived mathematical relation rather than a self-referential fit, and the manuscript treats Tweedie's formula and Fisher information as external inputs.
Axiom & Free-Parameter Ledger
axioms (2)
- domain assumption Conditional densities belong to natural exponential families for the exact Bayesian-score-driven equivalence
- domain assumption Fisher scoring yields a valid local Gaussian approximation to the posterior correction
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.