Pith. sign in

REVIEW 4 major objections 5 minor 44 references

Bayesian Forecast Combination with Predictive Priors via Particle Filtering

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A Bayesian combination method that feeds model diversity into the weight process outperforms standard time-varying weighting in oil and macro forecasts.

desk verdict The DTVW idea is a legitimate extension of TVW, but the empirical gains are not credible because the CRPS grid search for initial parameters is run on the evaluation sample. read the letter →

arxiv 2508.07136 v2 pith:ST4UUKND submitted 2025-08-10 stat.ME

Xiaorui Luo , Yanfei Kang , Xue Luo This is my paper · ORCID
classification stat.ME MSC 62F1562M2062P20
keywords Bayesianforecastcombinationpredictivepriorsmodeldiversitytime-varyingweightsparticlefilteringdensityforecastsoilpriceforecastingmacroeconomic
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper proposes a Bayesian forecast combination framework that puts forward-looking information directly into the time-varying weights that average individual model forecasts. It instantiates the idea with model diversity: the scaled pairwise distances among constituent models' multi-step predictions enter the latent weight process as a predictive prior, penalizing redundant models and favoring complementary ones. The resulting diversity-driven time-varying weights (DTVW) are estimated by particle filtering and compared against equal weighting, Bayesian model averaging, and standard time-varying weights. In simulations and in oil-price and U.S. macro applications, DTVW reports the lowest RMSFE, log score, and CRPS in nearly every setting, with the largest gains appearing in longer-horizon density forecasts.

What carries the argument

The load-bearing object is the diversity-augmented latent weight process: a first-order Markov regression for the logit-scale weights $x_t$ that adds the scaled diversity vector $\mathrm{div}_{t,h}$ as a regressor. Scaled diversity $div^l_{k,t,h} = \sum_i(\tilde y^l_{k,t+h}-\tilde y^l_{i,t+h})^2 / \sum_{i,j}(\tilde y^l_{i,t+h}-\tilde y^l_{j,t+h})^2$ measures how far each model's $h$-step-ahead prediction sits from the others, normalized to be scale-free. The coefficients $\theta_t$ balancing history against diversity are themselves latent and random-walking via $\alpha_t$, with final weights obtained by the softmax of $x_t$; the particle filter updates this system as new data arrive. This ma

What would settle it

Compute the two-stage grid search for $(\alpha_{1,0},\alpha_{2,0})$ using only data up to, say, 1991:12 for the oil application, fix those initial values, then re-run the out-of-sample evaluation from 1992:06 onward. If DTVW no longer beats TVW on the 1-, 3-, and 6-step metrics of Table 3, the reported edge is a tuning artifact rather than a property of the diversity-augmented weight process.

Watch

Extended reading notes

Core claim

The paper claims that forecast combination weights should be learned from future model behavior, not only from past forecast errors. The proposed DTVW model replaces the random-walk weight dynamic of standard TVW with the regression $x_t = \theta_{0,t} + \theta_{1,t} x_{t-1} + \theta_{2,t}\,\mathrm{div}_{t,h} + \varepsilon_{1,t}$, where $\mathrm{div}_{t,h}$ is a scaled diversity of the $h$-step-ahead predictions of the $K$ candidate models. The coefficients $\theta_t$ live on the cube $[-1,1]^3$ through a logistic link from a Brownian latent process, and all unknowns are propagated by a particle filter. The paper reports that the estimated coefficient on diversity is consistently positive ac

Load-bearing premise

The reported DTVW gains assume that the continuous-ranked-probability-score grid search for initial values $(\alpha_{1,0},\alpha_{2,0})$ is run on information available before the out-of-sample evaluation window; if the same data are used for tuning and scoring, the comparison favors DTVW by construction.

Editorial extensions

If this is right

  • If the reported results hold, forecast combinations can become anticipatory: weights begin shifting toward models whose future predictions diverge before their relative accuracy is confirmed by realized outcomes.
  • The diversity term provides a practical diagnostic: a persistently positive $\theta_{2,t}$ signals that the model set is misspecified or incomplete, while in a complete model set DTVW reverts toward TVW and still converges to the true model.
  • Longer-horizon density forecasts, where the paper reports the largest improvements, stand to benefit most, because multi-step predictions carry the diversity signal further ahead.
  • The framework is agnostic to the forward-looking signal; replacing diversity with another anticipatory measure only changes the regressor in the latent process.
  • In bivariate applications, tuning the initial coefficients for one variable (PCE) can slightly degrade the other variable's CRPS, a trade-off made visible through the joint particle-filter update.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A natural out-of-sample validation would redo the two-stage grid search on a training window only; the paper does not state that the CRPS used to pick $(\alpha_{1,0},\alpha_{2,0})$ is computed outside the evaluation period, so a skeptical reader should check this first.
  • The same regressor could be replaced by forecast disagreement, scoring-rule gaps, or volatility indices; if DTVW's advantage survives a placebo test with randomized $\mathrm{div}_{t,h}$, the diversity signal itself is doing causal work.
  • The negative estimates of $\alpha_{1,t}$ on real data, combined with positive $\alpha_{2,t}$, suggest the model learns to distrust historical extrapolation in volatile markets; an extension could make $\theta_{1,t}$ respond to volatility regimes in a more structured way than a random walk.
  • Because the diversity signal uses current multi-step predictions of candidate models, the method is not fully real-time if those predictions are revised ex post; one could test sensitivity to using only vintage forecasts.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes a Bayesian forecast combination method, DTVW, that extends the TVW approach of Billio et al. (2013) by adding a scaled model-diversity regressor div_{t,h} to the latent weight evolution equation. The latent coefficients are modeled as logistic transforms of a Brownian motion and estimated by particle filtering. The authors report simulation evidence, an oil price forecasting application, and a bivariate U.S. inflation/GDP application, concluding that DTVW consistently outperforms TVW and other benchmarks. The central empirical claim is that the diversity regressor produces real out-of-sample gains.

Significance. If validated, the proposed framework would be a useful extension of dynamic Bayesian forecast combination: it is simple, builds directly on a well-known method, and the particle-filter implementation is clearly specified. The simulation designs are fully described, and the macroeconomic application uses the same data as Billio et al. (2013), which aids comparability. However, the central empirical claim is currently not established because the initial values (alpha1,0, alpha2,0) are selected by minimizing CRPS on what appears to be the same data subsequently used for forecast evaluation. This is a load-bearing issue: the reported gains in Tables 1-4 and the claimed recovery of TVW in Section 4.2.1 may be tuning artifacts rather than evidence for the diversity mechanism.

major comments (4)
  1. [Sections 4.2.1-5.2; Figures 3, 6, 9, 14; Tables 1-4] The grid search for (alpha1,0, alpha2,0) in the DTVW is selected by minimizing CRPS, but no validation split disjoint from the evaluation period is ever described. The oil evaluation runs over 1992:06-2024:08 and the macro evaluation over 1970:Q1-2009:Q4, and the two-stage grid searches appear to use those same windows. Since Section 4.2.1 states that initialization significantly affects model performance, the reported DTVW improvements over TVW in Tables 1-4 may reflect test-set fitting. Please provide a nested or rolling validation scheme, or demonstrate that the ordering of results is robust to choosing initial values on a training/validation period only.
  2. [Section 5.3; Figures 17-19] The experiment that sets alpha1,0 = alpha2,0 = 0 is presented as evidence that the parameters are determined by the data itself rather than initial parameter values, but only the estimated parameter trajectories are shown. No forecast accuracy metrics (RMSFE, LS, CRPS) are reported for the zero-initialization DTVW. This does not address whether the out-of-sample gains in Tables 3-4 survive without tuning, so it cannot rescue the main claim. Please report the evaluation metrics for zero-initialized DTVW against TVW in all applications.
  3. [Section 3.2.1, Eq. (3.8)] The label predictive prior is overstated. In Eq. (3.8), div_{t,h} is a deterministic function of the candidate models' h-step forecasts and is used as a time-varying covariate in the state equation for x_t. It is not a prior distribution over future outcomes or over parameters, in the sense of the predictive priors literature cited in the introduction. This does not invalidate the mechanism, but the abstract and Section 6 should be reworded to describe forward-looking covariates rather than predictive priors.
  4. [Section 3.2.3, Algorithm 1; Eqs. (3.15)-(3.16)] The covariance matrices Sigma, Sigma1,t, Sigma2,t, and Lambda are never specified. If they are fixed tuning constants, their values should be stated; if they are estimated online, the priors/updating scheme should be provided. If any of these are also tuned on the evaluation sample, the test-set fitting concern of the first major comment applies to them as well.
minor comments (5)
  1. [Section 2.3] The BMA weight formula has mangled summation indices: 't sum t=t' should be the beginning and end of the evaluation period, with the range clearly defined. This makes the definition hard to read.
  2. [Section 5.1; Figure 10] The text says forecasts start in 1992:01, but Figure 10 labels the evaluation period as 1992:06-2024:08. Please clarify how many initial forecasts are discarded and use consistent dates.
  3. [Section 4.1] The sentence 'see Section 4.2 (Gneiting and Raftery, 2007)' should be a citation to Gneiting and Raftery (2007), not a cross-reference to Section 4.2.
  4. [Section 2.2, Eq. (2.12)] The scaled diversity is undefined when all forecasts in the denominator are equal (division by zero). Please state the convention used in degenerate cases.
  5. [Section 3.2.1, Eq. (3.9)] The text says theta_t is in [-1,1]^3, but the logistic transformation in Eq. (3.9) maps alpha_i,t to the open interval (-1,1). This is a minor notational imprecision; the text should say (-1,1) or allow endpoints by convention.

Circularity Check

2 steps flagged · score 6.0 of 10

Grid-searched initialization on evaluation-period CRPS makes DTVW's reported gains partly tuning artifacts; the adaptive-TVW 'recovery' is also selection-driven.

  1. fitted input called prediction [Section 5.1, Figures 9–10, Table 3; also Sections 4.2.1–4.2.2 and 5.2]
    "As illustrated in Section 4.2, the initialization of the forecast parameters in a step ahead (α1,0, α2,0) in DTVW is crucial to the performance. In this section, we perform the two-stage grid search as in the complex incomplete model set (4.3); see Section 4.2.2. ... The parameters with the lowest CRPS are located at (α1,0, α2,0) = (−3, 4.5)."

    The DTVW's headline advantage over TVW is measured on the oil forecast evaluation period 1992:06–2024:08 (Figure 10 caption, Table 3). The same CRPS criterion is used to select the initial latent parameters (α1,0, α2,0), and no training/validation subsample disjoint from the evaluation period is stated. Since the paper itself says initialization 'significantly affects model performance,' the reported 11–18% improvements over TVW are at least partly produced by minimizing the very score used to declare DTVW superior. The same pattern appears in the simulations (Figures 3 and 6 vs Tables 1–2) and in the macro application (Figure 14 vs Table 4).

  2. fitted input called prediction [Section 4.2.1, Figures 3–4, Table 1]
    "The left panel shows the CRPS decreases dramatically as α1,0 increasing from 0 to 10. It suggests that the minimal CRPS is achieved at α1,0≈ 9, i.e. θ1,0≈ 0.9998, which is extremely close to 1. ... This verifies numerically that our DTVW by setting α2,t≡ θ2,t≡ 0 and properly chosen the initialization of (α0,0, α1,0), can recover the TVW ... This adaptive TVW gives a data-driven explanation to why θ0,0 = 0, θ1,0 = 1 is chosen in TVW."

    The claimed 'recovery' of the TVW restriction (θ0=0, θ1=1) is not an independent finding: the adaptive TVW's initial α1,0 is chosen by grid search to minimize CRPS on the same simulated paths later scored in Table 1, and that CRPS-minimizing value is θ1≈1. The posterior path then remains near the selected initial value. Thus the 'data-driven explanation' reduces to the selection rule: a fitted initial parameter is relabeled as evidence for the TVW calibration.

full rationale

The core state-space construction is not circular: Eq. (3.8) augments the latent weight process with the diversity regressor div_{t,h}, which is computed from ex ante multi-step forecasts (Eq. 2.12), not from the outcomes being predicted. The particle-filter implementation (Algorithm 1) and the Bayesian updating equations are self-contained. The label 'predictive prior' is loose—div_{t,h} enters as a time-varying covariate in the state equation rather than as a prior over future outcomes—but that is a naming/conceptual issue, not a definitional reduction of the derivation to its inputs. The material circularity is in the empirical evaluation: DTVW has extra tunable initial parameters, and the paper selects them by minimizing CRPS without documenting a disjoint validation window, then reports out-of-sample superiority on the same evaluation period. The adaptive-TVW recovery of θ1≈1 is likewise an artifact of selecting α1,0 by CRPS. These are instances of fitted inputs being presented as predictive findings, so the headline empirical claims are partially forced. No load-bearing self-citation or imported uniqueness theorem is present; the diversity metric itself is defined in the paper. Score 6 reflects partial circularity confined to the tuning/evaluation chain, while the model formulation retains independent content.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The method rests on standard Bayesian state-space machinery plus several paper-specific choices: the scaled diversity signal, the logistic scaling of coefficients, the Gaussian assumptions, and critically the grid-selected initial latent values. The most consequential entries are the fitted initial parameters α1,0 and α2,0 and the unspecified covariance matrices, because the reported gains are sensitive to them.

free parameters (4)
  • α1,0, initial persistence coefficient = simple sim: 9 or 10; complex sim: 7; oil: -3; macro: -2
    Selected by two-stage grid search to minimize CRPS; the paper says initialization significantly affects model performance, so these are tuned free parameters.
  • α2,0, initial diversity coefficient = simple sim: 8.5; complex sim: 7; oil: 4.5; macro: 9
    Selected by grid search to minimize CRPS on the same datasets used for evaluation; not derived from theory.
  • α0,0, initial intercept = 0
    Set to zero by convention; affects θ0,0 through the logistic scaling in Equation (3.9).
  • Covariance matrices Σ, Σ1,t, Σ2,t, Λ = unspecified
    Required by Equations (2.9), (2.10), (3.15), and (3.16) for the likelihood and state transitions, but their values or priors are never given. They are effectively additional tuning choices.
assumptions (6)
  • standard math Gaussian likelihood and Gaussian transition densities for the latent weight process (Equations 2.9, 3.15).
    The paper adopts Gaussian forms from Billio et al. (2013) to make the particle filter tractable.
  • standard math Softmax logistic link maps latent states to simplex-valued combination weights (Equation 2.11).
    Used to guarantee weights are nonnegative and sum to one.
  • domain assumption At time t-1, h-step-ahead forecasts from all K models are available and can be used to form the prior for Wt without look-ahead (Assumption (0), Equations 3.1-3.2).
    The entire forward-looking feedback framework rests on this timing assumption; if the diversity term uses information not available in real time, the method is not a valid forecasting procedure.
  • domain assumption The prior density of Wt depends on forward forecasts up to t+h-1 but not beyond (Assumption (2)', Equation 3.5).
    This truncation is introduced ad hoc to justify using diversity as a predictive prior.
  • ad hoc to paper The scaled diversity measure in Equation (2.12) is an appropriate forward-looking signal for weight adjustment.
    The paper borrows the diversity concept from Kang et al. (2022) but provides no optimality or completeness argument for this particular function of the forecasts.
  • ad hoc to paper The latent coefficients θt are linked to a Brownian process αt via a logistic scaling to [-1,1] (Equations 3.9-3.10).
    This scaling is chosen for comparability with the diversity measure, not derived from any forecasting theory.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Bayesian Forecast Combination with Predictive Priors via Particle Filtering." pith.science (2026). https://pith.science/paper/ST4UUKND

@misc{pith2026250807136,
  author       = {Pith},
  title        = {Pith review of: Bayesian Forecast Combination with Predictive Priors via Particle Filtering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ST4UUKND}},
  note         = {Machine review of arXiv:2508.07136}
}
read the original abstract

We propose a Bayesian forecast combination framework that, for the first time, embeds forward-looking signals, formulated as predictive priors, directly into the time-varying weight-updating process. This approach enables weights to adapt using both historical forecast performance and anticipated future model behavior. We implement the framework with model diversity as the forward-looking signal, yielding the diversity-driven time-varying weights (DTVW) method. Compared with the standard time-varying weights (TVW) approach, DTVW embeds diversity-driven predictive priors that penalize redundancy and encourage informative contributions across constituent models. Simulation experiments, covering both a simple complete model set and a complex misspecified environment, show that DTVW improves forecast accuracy by dynamically focusing on well-performing models. Empirical applications to multi-step-ahead oil price forecasts and bivariate forecasts of U.S. inflation and GDP growth confirm its superiority over benchmarks including Equal weighting, Bayesian Model Averaging, and standard TVW. Beyond accuracy gains, diversity-based predictive priors provide diagnostic insights into model incompleteness and forecast uncertainty, making DTVW both more adaptive and more informative than existing Bayesian combination methods.

Figures

Figures reproduced from arXiv: 2508.07136 by the authors.

Figure 2
Figure 2. The blue fonts therein emphasize the differences from tho [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

44 extracted references · 41 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in ":" * " " * FUNCTION f...

  2. [2]

    , author Cross, J.L

    author Aastveit, K.A. , author Cross, J.L. , author van Dijk, H.K. , year 2023 . title Quantifying time-varying forecast uncertainty and risk for the real price of oil . journal Journal of Business & Economic Statistics volume 41 , pages 523--537

    work page 2023

  3. [3]

    , author Mitchell, J

    author Aastveit, K.A. , author Mitchell, J. , author Ravazzolo, F. , author van Dijk, H.K. , year 2019 . title The evolution of forecast density combinations in economics . journal Oxford Research Encyclopedia of Economics and Finance volume 4 , pages 1--39

    work page 2019

  4. [4]

    , author Giacomini, R

    author Amisano, G. , author Giacomini, R. , year 2007 . title Comparing density foreasts via weighted likelihood ratio tests . journal Journal of Business & Economic Statistics volume 25 , pages 177--190

    work page 2007

  5. [5]

    , author Dua, P

    author Batchelor, R. , author Dua, P. , year 1995 . title Forecaster diversity and the benefits of combining forecasts . journal Management Science volume 41 , pages 68--75

    work page 1995

  6. [6]

    , author Granger, C.W.J

    author Bates, J.M. , author Granger, C.W.J. , year 1969 . title The combination of forecasts . journal Journal of the Operational Research Society volume 20 , pages 451--468

    work page 1969

  7. [7]

    , author Casarin, R

    author Bernardi, M. , author Casarin, R. , author Maillet, B.B. , author Petrella, L. , year 2024 . title Bayesian dynamic quantile model averaging . journal Annals of Operations Research :https://doi.org/10.1007/s10479-024-06378-7. note online first

    work page doi:10.1007/s10479-024-06378-7 2024

  8. [8]

    , author Casarin, R

    author Billio, M. , author Casarin, R. , author Ravazzolo, F. , author van Dijk, H.K. , year 2013 . title Time-varying combinations of predictive densities using nonlinear filtering . journal Journal of Econometrics volume 177 , pages 213--232

    work page 2013

Show all 44 references
  1. [9]

    , author Grassi, S

    author Casarin, R. , author Grassi, S. , author Ravazzolo, F. , author van Dijk, H.K. , year 2015 . title Parallel sequential M onte C arlo for efficient density combination: The DECO M atlab toolbox . journal Journal of Statistical Software volume 68 , pages 1--30

    2015

  2. [10]

    , author Pauwels, L.L

    author Chan, F. , author Pauwels, L.L. , year 2018 . title Some theoretical results on forecast combinations . journal International Journal of Forecasting volume 34 , pages 64–74

    2018

  3. [11]

    , author Hsiao, C.Y

    author Chan, J.C. , author Hsiao, C.Y. , year 2014 . title Estimation of stochastic volatility models with heavy tails and serial dependence , in: editor Jeliazkov, I. , editor Yang, X.S. (Eds.), booktitle Bayesian Inference in the Social Sciences . publisher John Wiley & Sons...

    2014

  4. [12]

    , author Magnus, J.R

    author Claeskens, G. , author Magnus, J.R. , author Vasnev, A.L. , author Wang, W. , year 2016 . title The forecast combination puzzle: A simple theoretical explanation . journal International Journal of Forecasting volume 32 , pages 754–762

    2016

  5. [13]

    , year 1989

    author Clemen, R.T. , year 1989 . title Combining forecasts: A review and annotated bibliography . journal International Journal of Forecasting volume 5 , pages 559--583

    1989

  6. [14]

    , author Hasegawa, B.R

    author Del Negro, M. , author Hasegawa, B.R. , author Schorfheide, F. , year 2016 . title Dynamic prediction pools: An investigation of financial frictions and forecasting performance . journal Journal of Econometrics volume 192 , pages 391--405

    2016

  7. [15]

    , author Mariano, R.S

    author Diebold, F.X. , author Mariano, R.S. , year 1995 . title Comparing predictive accuracy . journal Journal of Business & Economic Statistics volume 13 , pages 253--263

    1995

  8. [16]

    , author Freitas, J.G

    author Doucet, A. , author Freitas, J.G. , author Gordon, J. , year 2001 . title Sequential M onte C arlo methods in practice . publisher Springer-Verlag , address New York

    2001

  9. [17]

    , author Bouaynaya, N

    author Ebinger, B. , author Bouaynaya, N. , author Polikar, R. , author Shterenberg, R. , year 2015 . title Constrained state estimation in particle filters , in: booktitle 2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pp. pages 4050--4054

    2015

  10. [18]

    , author Simpson, D

    author Gabry, J. , author Simpson, D. , author Vehtari, A. , author Betancourt, M. , author Gelman, A. , year 2019 . title Visualization in bayesian workflow . journal Journal of the Royal Statistical Society Series A: Statistics in Society volume 182 , pages 389--402

    2019

  11. [19]

    , author Vahey, S.P

    author Garratt, A. , author Vahey, S.P. , author Zhang, Y. , year 2019 . title Real-time forecast combinations for the oil price . journal Journal of Applied Econometrics volume 34 , pages 456--462

    2019

  12. [20]

    , author Simpson, D

    author Gelman, A. , author Simpson, D. , author Betancourt, M. , year 2017 . title The prior can often only be understood in the context of the likelihood . journal Entropy volume 19 , pages 555

    2017

  13. [21]

    , author Vehtari, A

    author Gelman, A. , author Vehtari, A. , author Simpson, D. , author Margossian, C.C. , author Carpenter, B. , author Yao, Y. , author Kennedy, L. , author Gabry, J. , author B \"u rkner, P.C. , author Modr \'a k, M. , year 2020 . title Bayesian workflow . journal arXiv prepri...

    2020 arXiv

  14. [22]

    , author Schervish, M.J

    author Genest, C. , author Schervish, M.J. , year 1985 . title Modelling expert judgements for bayesian updating . journal The Annals of Statistics volume 13 , pages 1198--1212

    1985

  15. [23]

    , author Vasnev, A.L

    author Gibbs, C.G. , author Vasnev, A.L. , year 2024 . title Conditionally optimal weights and forward-looking approaches to combining forecasts . journal International Journal of Forecasting volume 40 , pages 1734--1751

    2024

  16. [24]

    , author Raftery, A

    author Gneiting, T. , author Raftery, A. , year 2007 . title Strictly proper scoring rules, prediction, and estimation . journal Journal of the American Statistical Association volume 102 , pages 359--378

    2007

  17. [25]

    , author Ranjan, R

    author Gneiting, T. , author Ranjan, R. , year 2013 . title Combining predictive distributions . journal Electronic Journal of Statistics. volume 7 , pages 1747--1782

    2013

  18. [26]

    , year 2023

    author Graefe, A. , year 2023 . title Embrace the differences: Revisiting the PollyVote method of combining forecasts for U.S. presidential elections (2004 to 2020) . journal International Journal of Forecasting volume 39 , pages 170--177

    2023

  19. [27]

    , author Mitchell, J

    author Hall, S.G. , author Mitchell, J. , year 2007 . title Combining density forecasts . journal International Journal of Forecasting volume 23 , pages 1--13

    2007

  20. [28]

    , author Agiashvili, G

    author Hartmann, M. , author Agiashvili, G. , author B\" u rkner, P. , author Klami, A. , year 2020 . title Flexible prior elicitation via the prior predictive distribution , in: booktitle Proceedings of the 36th Conference on Uncertainty in Artificial Intelligence ( UAI ) , p...

    2020

  21. [29]

    , author Madigan, D

    author Hoeting, J.A. , author Madigan, D. , author Raftery, A.E. , author Volinsky, C.T. , year 1999 . title Bayesian model averaging: A tutorial . journal Statistical Science volume 14 , pages 382--401

    1999

  22. [30]

    , author Mitchell, J

    author Jore, A.S. , author Mitchell, J. , author Vahey, S.P. , year 2010 . title Combining forecast densities from VAR s with uncertain instabilities . journal Journal of Applied Econometrics volume 25 , pages 621--634

    2010

  23. [31]

    , author Cao, W

    author Kang, Y. , author Cao, W. , author Petropoulos, F. , author Li, F. , year 2022 . title Forecast with forecasts: Diversity matters . journal European Journal of Operational Research volume 301 , pages 180--190

    2022

  24. [32]

    , author Winkler, R.L

    author Lichtendahl, K.C. , author Winkler, R.L. , year 2020 . title Why do some combinations perform better than others? journal International Journal of Forecasting volume 36 , pages 142--149

    2020

  25. [33]

    , author Aastveit, K.A

    author McAlinn, K. , author Aastveit, K.A. , author Nakajima, J. , author West, M. , year 2020 . title Multivariate bayesian predictive synthesis in macroeconomic forecasting . journal Journal of the American Statistical Association volume 115 , pages 1092--1110

    2020

  26. [34]

    , author West, M

    author McAlinn, K. , author West, M. , year 2019 . title Dynamic B ayesian predictive synthesis in time series forecasting . journal Journal of Econometrics volume 210 , pages 155--169

    2019

  27. [35]

    , author Michaelson, K

    author Pogorelsky, B. , author Michaelson, K. , author Zanetti, R. , year 2022 . title Particle filter with lmmse importance sampling , in: booktitle 2022 25th International Conference on Information Fusion (FUSION) , organization IEEE . pp. pages 1--6

    2022

  28. [36]

    , author Gneiting, T

    author Raftery, A.E. , author Gneiting, T. , author Balabdaoui, F. , author Polakowski, M. , year 2005 . title Using bayesian model averaging to calibrate forecast ensembles . journal Monthly Weather Review volume 133 , pages 1155--1174

    2005

  29. [37]

    , author K \'a rn \'y , M

    author Raftery, A.E. , author K \'a rn \'y , M. , author Ettler, P. , year 2010 . title Online prediction under model uncertainty via dynamic model averaging: Application to a cold rolling mill . journal Technometrics volume 52 , pages 52--66

    2010

  30. [38]

    , author van Dijk, H.K

    author Terui, N. , author van Dijk, H.K. , year 2002 . title Combined forecasts from linear and nonlinear time series models . journal International Journal of Forecasting volume 18 , pages 421--438

    2002

  31. [39]

    , author Pollock, A.C

    author Thomson, M.E. , author Pollock, A.C. , author Önkal, D. , author Gönül, M.S. , year 2019 . title Combining forecasts: Performance and coherence . journal International Journal of Forecasting volume 35 , pages 474--484

    2019

  32. [40]

    , year 2006

    author Timmermann, A. , year 2006 . title Forecast combinations , in: editor Elliott, G. , editor Granger, C.W.J. , editor Timmermann, A. (Eds.), booktitle Handbook of Economic Forecasting . publisher North-Holland, Elsevier . volume volume 1 , pp. pages 135--196

    2006

  33. [41]

    , year 2025

    author Villani, M. , year 2025 . title Bayesian learning . https://github.com/mattiasvillani/BayesianLearningBook. note unpublished manuscript; first edition (February 2025)

    2025

  34. [42]

    , author Hyndman, R.J

    author Wang, X. , author Hyndman, R.J. , author Li, F. , author Kang, Y. , year 2023 . title Forecast combinations: An over 50-year review . journal International Journal of Forecasting volume 39 , pages 1518--1547

    2023

  35. [43]

    , year 1984

    author West, M. , year 1984 . title Bayesian aggregation . journal Journal of the Royal Statistical Society: Series A (General) volume 147 , pages 600--607

    1984

  36. [44]

    , author Crosse, J

    author West, M. , author Crosse, J. , year 1992 . title Modelling of probabilistic agent opinion . journal Journal of the Royal Statistical Society: Series B (Methodological) volume 54 , pages 285--299

    1992

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.