Some cautionary tales about Bayesian predictive inference

Emanuela Dreassi; Fabrizio Leisen; Luca Pratelli; Pietro Rigo

REVIEW 2 major objections 4 minor 18 references

Asymptotic exchangeability — the shifted tail of a data sequence converging in distribution to an exchangeable sequence — is far too weak to license treating the tail as approximately exchangeable or as inferentially equivalent to its excha

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 13:07 UTC pith:POPN32NF

load-bearing objection Asymptotic exchangeability is indeed too weak for inference — the warning is correct, but Example 7's unproved 'it can be shown' claims need proofs or references before this becomes a citable caution. the 2 major comments →

arxiv 2607.19206 v1 pith:POPN32NF submitted 2026-07-21 math.ST stat.MEstat.MLstat.TH

Some cautionary tales about Bayesian predictive inference

Emanuela Dreassi , Fabrizio Leisen , Luca Pratelli , Pietro Rigo This is my paper

classification math.ST stat.MEstat.MLstat.TH MSC 62F1562A9962G0962M2060G25

keywords asymptotic exchangeabilityBayesian predictive inferencepredictive distributionsexchangeable limit sequenceweak convergence of measureslaw of large numberscredible intervalstotal variation convergence

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Bayesian predictive inference often proceeds by specifying predictive distributions, then treats a data sequence as if, after enough observations, its unobserved tail is nearly exchangeable and therefore equivalent to an exchangeable limit sequence. This paper argues that asymptotic exchangeability — the shifted tail converging in distribution to some exchangeable sequence — does not license that reading. The sharpest construction has an exchangeable limit while the data's empirical mean fails to converge in probability, the data and the limit sequence have mutually singular distributions, and the data's predictive distributions fail to converge while the limit's do. A second line of examples shows that asymptotic credible intervals valid for the exchangeable limit need not be valid for the original data. If the paper is right, applied Bayesian work should require stronger types of predictive convergence before treating tails as nearly exchangeable.

Core claim

The paper's central claim is that asymptotic exchangeability, defined as convergence in distribution of (X_{n+1}, X_{n+2}, ...) to an exchangeable sequence Z, is far too weak to justify common Bayesian interpretations: that for large n the tail is approximately i.i.d. given some random measure, or that all sequences with the same exchangeable limit are inferentially equivalent. The main demonstration is a construction with X_n = n^{-1/2} (U_1 + ... + U_n) + V_n, where U and V are independent standard-normal i.i.d. sequences. Its shifted tail converges in distribution to Z = (U_1+V_1, U_1+V_2, ...), an exchangeable sequence, yet the paper argues that the empirical mean of X fails to converge

What carries the argument

The relation X∼Z, in which two data sequences share the same exchangeable limit sequence in distribution, is the object whose inferential weight is interrogated. The counterexample engine is the predictive construction X_n = n^{-1/2} sum_{i=1}^n U_i + V_n with independent standard-normal U and V sequences; its tail has the same distributional limit as Z = (U_1+V_1, U_1+V_2, ...), but its empirical mean, probability law, and predictive distributions behave differently from Z's. A supporting mechanism compares two predictive sequences that are asymptotically close in a probability metric and shows that they can still induce very different data laws.

Load-bearing premise

The argument's main demonstration rests on three assertions in Example 7 that the paper states with 'it can be shown' and does not prove: the empirical mean of the constructed X fails to converge in probability, X and the exchangeable limit Z have mutually singular distributions, and X's predictives fail to converge weakly; if any of these assertions is false, the claim that asymptotic exchangeability does not even imply the law of large numbers loses its key support.

What would settle it

Check the unproved part of Example 7 directly: compute or simulate whether Xbar_n = (1/n) sum_{i=1}^n (i^{-1/2} sum_{k=1}^i U_k + V_i) converges in probability. If it does converge, the paper's main demonstration fails; if it provably does not, the cautionary conclusion stands. The covariance calculation alone cannot settle the convergence question.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

Verifying that the data tail converges in distribution to an exchangeable sequence is not enough to justify Bayesian updates that assume near-exchangeability of the tail; a stronger mode of predictive convergence is needed.
The law of large numbers for the original data does not follow from asymptotic exchangeability: the paper's construction has an exchangeable limit while its empirical mean fails to converge in probability.
Asymptotic credible intervals computed from the exchangeable limit sequence can have wrong coverage for the original data, because the studentized CLT that holds for exchangeable sequences can fail along X.
Two predictive sequences that are asymptotically close in a metric can still generate qualitatively different data sequences, so closeness of predictives is not a license to substitute one model for another.
Stronger conditions, such as total-variation convergence of the predictives or of the shifted sequence law, would restore the approximate-exchangeability interpretation.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The covariance calculation suggests a practical diagnostic: before applying an exchangeable-limit approximation, estimate Cov(X_{n+i}, X_{n+j}) for several lags and check whether it stabilizes as n grows; the pattern of covariances drifting toward both 0 and 1 would warn against the approximation.
A testable extension is a simulation study that draws data from the paper's construction, fits the exchangeable limit model, and measures the actual coverage of nominal credible intervals, which would quantify the finite-sample gap.
The paper's strengthenings point to a program: characterize, for each inferential target (law of large numbers, central limit theorem, credible intervals), the minimal mode of predictive convergence required; total variation is sufficient but likely not necessary.
The two views of the data-generating mechanism imply a terminology fix for applied work: reserve 'predictive distribution' for the actual conditional law of the next observation given the past, and call estimators of an unknown sampling measure something else.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Desk Editor's Note

Asymptotic exchangeability is indeed too weak for inference — the warning is correct, but Example 7's unproved 'it can be shown' claims need proofs or references before this becomes a citable caution.

read the letter

Bottom line: the paper’s central warning is correct and worth taking seriously — asymptotic exchangeability is far too weak to justify treating the tail of a predictive sequence as if it were exchangeable. I checked the key math, and it holds. But the sharpest demonstration ships with several 'it can be shown' assertions and no proof, so the note isn’t yet self-contained enough to be a citable caution.

What’s actually new and good: the paper clearly separates two views of predictives — the first view, where predictives generate the data-generating law, and the second, where they are estimators of an unknown random measure. That distinction is already in the authors’ own [4], but the examples here (bootstrap, Hill’s model, Bayesian nonparametrics, predictive resampling) make it concrete and fair. The M2 warning is the real content: even if (X_{n+1}, X_{n+2}, …) converges in distribution to an exchangeable Z, the original sequence X can fail the SLLN, have singular distribution relative to Z, and have predictives that don’t converge while Z’s do. Example 7’s covariance computation Cov(X_{n+i}, X_{n+j}) = sqrt((n+i)/(n+j)) is simple and striking — it shows the tail is not even approximately exchangeable in second moments, for any n. I verified the SLLN failure and singularity claims: they are correct.

Soft spots: the 'it can be shown' claims in Example 7 are load-bearing. They are true, but the reader shouldn’t have to verify them independently. The authors should either prove them or cite the exact location in [16] where the argument is made. Example 8 depends entirely on the authors’ own [2] for the CLT failure; [2] is a published Annals paper, so that is acceptable, but an independent derivation would make the note stronger. The novelty is narrow — the M1 distinction is essentially from [4] and the counterexample model is 'as in Example 4.6 of [16]' — but that’s fine for a clarification note. No circularity, no overclaiming; the paper is honest and well-written. Self-citation here is not a problem because the cited results are real and published.

Who this is for: anyone using martingale posteriors, predictive resampling, Hill-model inference, or kernel-density posterior uncertainty — i.e., the papers [5], [6], [9], [12], [13] that treat the shifted sequence as approximately exchangeable. Those readers need this warning. It deserves peer review: a serious editor should send it out, with the expectation of a revision that fills the gaps in Example 7 (and ideally Example 8) before publication.

Referee Report

2 major / 4 minor

Summary. The paper is a short cautionary note on two misunderstandings in Bayesian predictive inference. The first (M1) is the conflation of two views of the data sequence: (i) P is constructed from a chosen sequence of predictive distributions via the Ionescu-Tulcea theorem, and (ii) P is an unknown structured law, with the chosen β_n serving as estimators of an unknown random measure. The second (M2) is the overinterpretation of asymptotic exchangeability: the belief that X∼Z (both shifted sequences converge to the same exchangeable limit) makes X and Z inferentially equivalent. The paper illustrates M1 with bootstrap, Hill's predictive distributions, Bayesian nonparametrics, and predictive resampling, and M2 with examples showing differences in SLLN behavior, CLT behavior, predictive convergence, and the effect of asymptotically equivalent predictives. It closes by proposing two possible strengthenings of asymptotic exchangeability.

Significance. If the mathematical claims are correct, the note fills a useful didactic gap: it crisply separates two meanings of 'predictive distribution' and gives a sharp warning that asymptotic exchangeability is too weak to license de Finetti-type conditional i.i.d. interpretations or asymptotic inferential equivalence. The examples are simple and readable; the covariance calculation in Example 7 is a transparent and correct demonstration. The paper makes no fitted-parameter claims and is honest about relying on published results, notably [2] and [16]. However, several headline claims in Example 7 are asserted without proof, and the final characterization of a total-variation strengthening is incorrect as stated. These issues are local and fixable, but they prevent the note from being citable in its current form.

major comments (2)

[§2, Example 7] The three load-bearing facts in Example 7 are each introduced by 'it can be shown' with no proof and only a vague reference to Example 4.6 of [16]: (i) X̄_n fails to converge in probability, (ii) X and Z have singular probability distributions, and (iii) the predictives of Z converge weakly while those of X do not. Fact (i) is the paper's demonstration that asymptotic exchangeability does not even imply the SLLN; fact (ii) is the sharpest evidence of inferential non-equivalence. These claims are not immediate from the displayed covariance calculation, which addresses tail exchangeability only. I have verified independently that the claims are mathematically correct, but a citable note should either prove them in an appendix or supply an exact theorem/example reference for each. As written, the central M2 conclusion rests on unverifiable assertions.
[§2, final remark (condition (ii))] The asserted equivalence for condition (ii) is false in general. Let Z be i.i.d. N(0,1) and set X_n = Z_n + a_n with a_n = 1/√n. Then X is asymptotically exchangeable with limit Z, and σ(X_{n+1},...) = σ(Z_{n+1},...) for every n, so the two laws agree on the trivial tail σ-field. Yet for each n the law of (X_{n+1}, X_{n+2}, ...) is an infinite product of N(a_{n+k},1) measures and is mutually singular with the law of Z by Kakutani's theorem, since Σ_k a_{n+k}^2 = ∞. Hence ∥ν_n - P_Z∥ = 1 for every n, not 0. The 'only if' direction is correct, but the converse needs additional hypotheses or should be withdrawn.

minor comments (4)

[§2, Example 9] The displayed formula for α_n^* is garbled; it should presumably read (c α_0 + n μ_n)/(n+c). Please fix.
[§2, Example 6] The abbreviation 'c.i.d.' is used without definition. Define it as 'conditionally identically distributed' and give the precise definition or a reference to [2].
[§2, Example 7] The covariance formula Cov(X_{n+i}, X_{n+j}) = √((n+i)/(n+j)) is correct, but a one-line derivation from the independence of U and V and the variance of partial sums would help the reader follow the point.
[§1.2] The definition of the relation 'Y∼X' uses the same symbol as the example 'X∼Z'; this is perhaps unavoidable, but consider an explicit phrase to avoid confusion with 'approximately distributed'.

Circularity Check

0 steps flagged

No circular derivation: the paper's examples are explicit constructions backed by published, parameter-free prior results; Example 7's unproved 'it can be shown' claims are a verifiability gap, not a circular reduction.

full rationale

No circularity is present. The paper contains no fitted parameters, no data-fitting step, and no quantity that is fitted to a subset of data and then renamed as a prediction. The M2 thesis—that asymptotic exchangeability is too weak to license treating the tail as approximately exchangeable—is supported by explicit counterexamples, not by a definitional identification. The load-bearing citations, Theorem 2.2 and Example 4.6 of [16] and Theorem 3.1/Example 3.2 of [2], are published, peer-reviewed, parameter-free mathematical results with stated assumptions that do not include the present conclusion; under the review rules these are independent evidence and do not raise the circularity score. The main caveat is in Example 7, where the assertions 'Firstly, it can be shown that \(\bar X_n = \frac1n \sum_{i=1}^n X_i\) fails to converge in probability', 'X and Z have singular probability distributions', and 'the predictive distributions of Z converge (weakly a.s.) while those of X do not' are introduced by 'it can be shown' with no proof or precise reference. These are load-bearing for the M2 demonstration, and the manuscript should supply proofs or references before it is cited as self-contained. But this is an omitted-proof/verifiability gap, not circularity: the claims are direct properties of the constructed Gaussian sequence \(X_n = \sum_{i=1}^n U_i/\sqrt n + V_n\), not restatements of the definition of asymptotic exchangeability, and no equation in the paper reduces the conclusion to its inputs. No other circularity pattern (self-definition, fitted input called prediction, uniqueness imported from authors, ansatz smuggled via citation, renaming of a known result) occurs.

Axiom & Free-Parameter Ledger

2 free parameters · 6 axioms · 0 invented entities

The note introduces no fitted parameters (the two constants c and q are arbitrary construction knobs) and no new objects: μ, Z, α_n, and the predictive resampling / Hill predictives are all standard constructs from the cited literature. Its novel content is in the examples, which rest on standard theorems plus the authors' own published counterexamples.

free parameters (2)

c (Example 9 concentration constant) = arbitrary positive constant
Appears in α*_n = (c α_0 + n μ_n)/(n+c); chosen by hand for the construction, not fitted to data. The conclusion d(α_n, α*_n) → 0 holds for any c > 0.
q (Example 6 weighting constant) = arbitrary in (0,1)
Weights in α_n = q^n α_0 + (1−q) Σ_i q^{n−i} δ_{X_i}; a construction constant used to exhibit permutation-non-invariant predictives. Not fitted.

axioms (6)

standard math Ionescu-Tulcea theorem: a law on R^∞ is uniquely determined by arbitrary predictives (α_n)
Defines the 'first view' in Section 1.1 and underlies all examples, where P is the law induced by (α_n).
standard math de Finetti's representation theorem for exchangeable sequences
Used throughout for the 'second view': exchangeable X implies P(X∈·) = E[μ^∞(·)] for some r.p.m. μ.
standard math Aldous Lemma (8.2): α_n → μ weakly a.s. implies asymptotic exchangeability
Invoked in Section 1.2 to connect predictive convergence to the exchangeable limit sequence Z.
standard math Published self-cited theorems: Theorem 2.2 of [16] (weak convergence in probability suffices), Theorem 3.1 of [2] (stable CLT for exchangeable Z), Example 3.2 of [2] (CLT failure for c.i.d. X)
These supply the non-exchangeable-X machinery; they are peer-reviewed, parameter-free results, treated as independent support though self-cited.
domain assumption A.s. existence of θ = lim_n (1/n)Σ_{i≤n} X_i
Explicitly stated as an assumption in Example 8: 'where the a.s. existence of the limit is an assumption'.
domain assumption A common probability space (Ω, A, P) carrying all sequences
Assumed in the notation section; required for a.s. statements about α_n and the empirical averages.

pith-pipeline@v1.3.0-alltime-deepseek · 7240 in / 19396 out tokens · 205363 ms · 2026-08-01T13:07:52.455021+00:00 · methodology

0 comments

read the original abstract

Two misunderstandings, frequently arising in Bayesian predictive inference, are discussed. The first deals with the data generating mechanism, while the second consists in overestimating the role played by asymptotic exchangeability. Some consequences of such misunderstandings are highlighted through examples.

discussion (0)

Reference graph

Works this paper leans on

18 extracted references · 1 linked inside Pith

[1]

(1985) Exchangeability and related topics,Ecole d’et´ e de Probabilit´ es de Saint- Flour XIII, Lecture Notes in Math

Aldous D.J. (1985) Exchangeability and related topics,Ecole d’et´ e de Probabilit´ es de Saint- Flour XIII, Lecture Notes in Math. 1117, Springer

1985
[2]

(2004) Limit theorems for a class of identically distributed random variables,Ann

Berti P., Pratelli L., Rigo P. (2004) Limit theorems for a class of identically distributed random variables,Ann. Probab., 32, 2029-2052

2004
[3]

(2013) Exchangeable sequences driven by an absolutely contin- uous random measure,Ann

Berti P., Pratelli L., Rigo P. (2013) Exchangeable sequences driven by an absolutely contin- uous random measure,Ann. Probab., 41, 2090-2102

2013
[4]

(2025) A probabilistic view on predictive constructions for Bayesian learning,Statist

Berti P., Dreassi E., Leisen F., Pratelli L., Rigo P. (2025) A probabilistic view on predictive constructions for Bayesian learning,Statist. Sci., 40, 25-39

2025
[5]

(2026) Posterior Inference via Hill’s Prediction Model, arXiv:2603.20071v1 8 EMANUELA DREASSI, F ABRIZIO LEISEN, LUCA PRATELLI, AND PIETRO RIGO

Bissiri P.G., Holmes C., Walker S.G. (2026) Posterior Inference via Hill’s Prediction Model, arXiv:2603.20071v1 8 EMANUELA DREASSI, F ABRIZIO LEISEN, LUCA PRATELLI, AND PIETRO RIGO

arXiv 2026
[6]

(2026) Posterior uncertainty for kernel density estimates, arXiv:2607.03927v1

Christensen D., Svardal T., Rønneberg L., Stoltenberg E.A. (2026) Posterior uncertainty for kernel density estimates, arXiv:2607.03927v1

Pith/arXiv arXiv 2026
[7]

(1999) Prequential probability: principles and properties,Bernoulli, 5, 125-162

Dawid A.P., Vovk V.G. (1999) Prequential probability: principles and properties,Bernoulli, 5, 125-162

1999
[8]

(2023) Neural diffusion processes,Proc

Dutordoir V., Saul A., Ghahramani Z., Simpson F. (2023) Neural diffusion processes,Proc. 40th Intern. Conf. on Machine Learning, PMLR 202, 8990-9012

2023
[9]

(2023) Martingale posterior distributions (with discussion), J

Fong E., Holmes C., Walker S.G. (2023) Martingale posterior distributions (with discussion), J. Royal Stat. Soc. B, 85, 1357-1391

2023
[10]

(2000) Exchangeability, predictive distributions and para- metric models,Sankhya A, 62, 86–109

Fortini S., Ladelli L., Regazzini E. (2000) Exchangeability, predictive distributions and para- metric models,Sankhya A, 62, 86–109

2000
[11]

(2012) Predictive construction of priors in Bayesian nonparametrics, Brazilian J

Fortini S., Petrone S. (2012) Predictive construction of priors in Bayesian nonparametrics, Brazilian J. Prob. Stat., 26, 423-449

2012
[12]

(2020) Quasi-Bayes properties of a procedure for sequential learning in mixture models,J

Fortini S., Petrone S. (2020) Quasi-Bayes properties of a procedure for sequential learning in mixture models,J. Royal Stat. Soc. B, 82, 1087-1114

2020
[13]

(2025) Exchangeability, prediction and predictive modeling in Bayesian statistics,Statist

Fortini S., Petrone S. (2025) Exchangeability, prediction and predictive modeling in Bayesian statistics,Statist. Sci., 40, 40–67

2025
[14]

(2018) On recursive Bayesian predictive distributions, J.A.S.A., 113, 1085-1093

Hahn P.R., Martin R., Walker S.G. (2018) On recursive Bayesian predictive distributions, J.A.S.A., 113, 1085-1093

2018
[15]

(1968) Posterior distribution of percentiles: Bayes’ theorem for sampling from a population,J.A.S.A., 63, 677–691

Hill B.M. (1968) Posterior distribution of percentiles: Bayes’ theorem for sampling from a population,J.A.S.A., 63, 677–691

1968
[16]

(2026) Weak convergence in probability of predictive distribu- tions,Electronic Commun

Leisen F., Pratelli L., Rigo P. (2026) Weak convergence in probability of predictive distribu- tions,Electronic Commun. Probab., 31, 1-15

2026
[17]

(2023) Bayesian causal inference: A critical review,Phil

Li F., Ding P., Mealli F. (2023) Bayesian causal inference: A critical review,Phil. Trans. R. Soc.A, 381

2023
[18]

G. Par- enti

Pitman J. (1996) Some developments of the Blackwell-MacQueen urn scheme,Statistics, Probability and Game Theory, IMS Lect. Notes Mon. Series, 30, 245-267. Emanuela Dreassi, Dipartimento di Statistica, Informatica, Applicazioni “G. Par- enti”, Universit`a di Firenze, viale Morgagni 59, 50134 Firenze, Italy Email address:emanuela.dreassi@unifi.it F abrizio ...

1996

[1] [1]

(1985) Exchangeability and related topics,Ecole d’et´ e de Probabilit´ es de Saint- Flour XIII, Lecture Notes in Math

Aldous D.J. (1985) Exchangeability and related topics,Ecole d’et´ e de Probabilit´ es de Saint- Flour XIII, Lecture Notes in Math. 1117, Springer

1985

[2] [2]

(2004) Limit theorems for a class of identically distributed random variables,Ann

Berti P., Pratelli L., Rigo P. (2004) Limit theorems for a class of identically distributed random variables,Ann. Probab., 32, 2029-2052

2004

[3] [3]

(2013) Exchangeable sequences driven by an absolutely contin- uous random measure,Ann

Berti P., Pratelli L., Rigo P. (2013) Exchangeable sequences driven by an absolutely contin- uous random measure,Ann. Probab., 41, 2090-2102

2013

[4] [4]

(2025) A probabilistic view on predictive constructions for Bayesian learning,Statist

Berti P., Dreassi E., Leisen F., Pratelli L., Rigo P. (2025) A probabilistic view on predictive constructions for Bayesian learning,Statist. Sci., 40, 25-39

2025

[5] [5]

(2026) Posterior Inference via Hill’s Prediction Model, arXiv:2603.20071v1 8 EMANUELA DREASSI, F ABRIZIO LEISEN, LUCA PRATELLI, AND PIETRO RIGO

Bissiri P.G., Holmes C., Walker S.G. (2026) Posterior Inference via Hill’s Prediction Model, arXiv:2603.20071v1 8 EMANUELA DREASSI, F ABRIZIO LEISEN, LUCA PRATELLI, AND PIETRO RIGO

arXiv 2026

[6] [6]

(2026) Posterior uncertainty for kernel density estimates, arXiv:2607.03927v1

Christensen D., Svardal T., Rønneberg L., Stoltenberg E.A. (2026) Posterior uncertainty for kernel density estimates, arXiv:2607.03927v1

Pith/arXiv arXiv 2026

[7] [7]

(1999) Prequential probability: principles and properties,Bernoulli, 5, 125-162

Dawid A.P., Vovk V.G. (1999) Prequential probability: principles and properties,Bernoulli, 5, 125-162

1999

[8] [8]

(2023) Neural diffusion processes,Proc

Dutordoir V., Saul A., Ghahramani Z., Simpson F. (2023) Neural diffusion processes,Proc. 40th Intern. Conf. on Machine Learning, PMLR 202, 8990-9012

2023

[9] [9]

(2023) Martingale posterior distributions (with discussion), J

Fong E., Holmes C., Walker S.G. (2023) Martingale posterior distributions (with discussion), J. Royal Stat. Soc. B, 85, 1357-1391

2023

[10] [10]

(2000) Exchangeability, predictive distributions and para- metric models,Sankhya A, 62, 86–109

Fortini S., Ladelli L., Regazzini E. (2000) Exchangeability, predictive distributions and para- metric models,Sankhya A, 62, 86–109

2000

[11] [11]

(2012) Predictive construction of priors in Bayesian nonparametrics, Brazilian J

Fortini S., Petrone S. (2012) Predictive construction of priors in Bayesian nonparametrics, Brazilian J. Prob. Stat., 26, 423-449

2012

[12] [12]

(2020) Quasi-Bayes properties of a procedure for sequential learning in mixture models,J

Fortini S., Petrone S. (2020) Quasi-Bayes properties of a procedure for sequential learning in mixture models,J. Royal Stat. Soc. B, 82, 1087-1114

2020

[13] [13]

(2025) Exchangeability, prediction and predictive modeling in Bayesian statistics,Statist

Fortini S., Petrone S. (2025) Exchangeability, prediction and predictive modeling in Bayesian statistics,Statist. Sci., 40, 40–67

2025

[14] [14]

(2018) On recursive Bayesian predictive distributions, J.A.S.A., 113, 1085-1093

Hahn P.R., Martin R., Walker S.G. (2018) On recursive Bayesian predictive distributions, J.A.S.A., 113, 1085-1093

2018

[15] [15]

(1968) Posterior distribution of percentiles: Bayes’ theorem for sampling from a population,J.A.S.A., 63, 677–691

Hill B.M. (1968) Posterior distribution of percentiles: Bayes’ theorem for sampling from a population,J.A.S.A., 63, 677–691

1968

[16] [16]

(2026) Weak convergence in probability of predictive distribu- tions,Electronic Commun

Leisen F., Pratelli L., Rigo P. (2026) Weak convergence in probability of predictive distribu- tions,Electronic Commun. Probab., 31, 1-15

2026

[17] [17]

(2023) Bayesian causal inference: A critical review,Phil

Li F., Ding P., Mealli F. (2023) Bayesian causal inference: A critical review,Phil. Trans. R. Soc.A, 381

2023

[18] [18]

G. Par- enti

Pitman J. (1996) Some developments of the Blackwell-MacQueen urn scheme,Statistics, Probability and Game Theory, IMS Lect. Notes Mon. Series, 30, 245-267. Emanuela Dreassi, Dipartimento di Statistica, Informatica, Applicazioni “G. Par- enti”, Universit`a di Firenze, viale Morgagni 59, 50134 Firenze, Italy Email address:emanuela.dreassi@unifi.it F abrizio ...

1996