Dynamic Bayesian Predictive Synthesis

• Dynamic Bayesian Predictive Synthesis is a Bayesian framework that combines multiple predictive distributions using time-evolving calibration parameters.
• It uses state-space models and synthesis functions to account for dynamic biases, miscalibration, and interdependencies among forecasts.
• Practical applications span macroeconomic forecasting, network analysis, and portfolio optimization, often yielding improved accuracy over single models.

Dynamic Bayesian Predictive Synthesis (BPS) is a formally Bayesian framework for combining multiple predictive distributions from models, forecasters, or other “agents” into a single predictive distribution whose calibration parameters evolve over time. In its core representation, the reported agent densities are treated as data, latent agent states are drawn from those densities, and a user-specified synthesis function maps the latent states to the outcome distribution. The static synthesis identity was given a rigorous proof by Masuda and Irie, while the dynamic time-series framework was developed by McAlinn and West through state-space models that allow time-varying biases, miscalibration, inter-dependencies, and forecast-combination weights (Masuda et al., 2024, McAlinn et al., 2016).

1. Foundational representation

The fundamental BPS equation starts with a decision maker (DM) facing $K$ agent forecast densities $h_k(x_k)$ and an information set $\mathcal H=\{h_1,\dots,h_K\}$. The DM’s posterior predictive density for outcome $y$ is

$$p(y\mid \mathcal H)=\int \alpha(y\mid x_1,\dots,x_K)\prod_{k=1}^K h_k(x_k)\,dx_1\cdots dx_K,$$

where $\alpha(y\mid x_{1:K})$ is a proper density in $y$ for each $x_{1:K}$ (Masuda et al., 2024). In the agent-opinion analysis perspective, the densities $h_j$ are themselves data, the latent variables $x_{1:J}$ are “agent states,” and the synthesis function $h_k(x_k)$0 encodes how the DM believes these latent forecasts relate to the realized outcome (McAlinn et al., 2016).

Masuda and Irie show that the static formula follows from a partially specified prior in which only marginal means of agent-reported probabilities are fixed, together with a consistency condition requiring that marginalization reproduces the DM’s unconditioned prior and that extreme forecasts match Bayes-rule values. They give two proofs: a proof by marginalization using induction on $h_k(x_k)$1, and a proof via an artificial independent prior that shares the same marginal means (Masuda et al., 2024). The first proof moves from the linear-in-deviations form for discretized forecast probabilities to a discrete mixture over joint-extreme conditional probabilities, and then takes a fine-grid limit. The second proof uses iterated expectations under a convenient independent prior to recover the same discrete mixture, and again takes the grid limit to the integral form (Masuda et al., 2024).

A central structural point is the equivalence between two representations: a linear-deviation form and a discrete-mixture form. In the continuous limit, the mixture form shows that $h_k(x_k)$2 is the DM’s implied conditional density of $h_k(x_k)$3 when each agent is degenerate at $h_k(x_k)$4 (Masuda et al., 2024). The graphical interpretation is a two-stage generative model in which $h_k(x_k)$5 independently and then $h_k(x_k)$6. Masuda and Irie emphasize that the product $h_k(x_k)$7 in this representation does not require agent-forecast independence in the DM’s actual prior; independence enters only as a proof device in one argument and as a conditional feature of the generative representation (Masuda et al., 2024).

2. Dynamic state-space constructions

Dynamic BPS embeds the static kernel in a time-indexed state-space model. At time $h_k(x_k)$8, agents deliver predictive densities $h_k(x_k)$9, the synthesis function may depend on a latent state $\mathcal H=\{h_1,\dots,h_K\}$0, and $\mathcal H=\{h_1,\dots,h_K\}$1 evolves by a Markovian law such as a random walk or a dynamic linear model (Masuda et al., 2024). In general form, one writes

$\mathcal H=\{h_1,\dots,h_K\}$2

with latent states and parameters $\mathcal H=\{h_1,\dots,h_K\}$3 governing the synthesis function (Chernis et al., 2023).

The canonical specification is linear-Gaussian. McAlinn and West formulate the synthesis density as

$\mathcal H=\{h_1,\dots,h_K\}$4

with

$\mathcal H=\{h_1,\dots,h_K\}$5

and state evolution

$\mathcal H=\{h_1,\dots,h_K\}$6

often with $\mathcal H=\{h_1,\dots,h_K\}$7 so that the coefficients follow a random walk (McAlinn et al., 2016). In this form, the intercept absorbs residual common bias, the coefficients on the latent agent states encode dynamic calibration, and $\mathcal H=\{h_1,\dots,h_K\}$8 controls the rate at which synthesis weights adapt (McAlinn et al., 2016). A closely related DLM formulation writes $\mathcal H=\{h_1,\dots,h_K\}$9, where $y$0 are auxiliary draws from the component densities $y$1 and the synthesized one-step-ahead density is obtained by integrating over both $y$2 and $y$3 (Kato et al., 8 Oct 2025).

Several named special cases appear in the recent literature. In the tree-based synthesis paper, the common parametric choice is a time-varying-parameter linear regression

$y$4

where $y$5 is a time-varying intercept, $y$6 are static combination weights, and $y$7 are dynamic weight adjustments. The variants BPS-CONST and BPS-RW respectively set $y$8 for all $y$9 or evolve $$p(y\mid \mathcal H)=\int \alpha(y\mid x_1,\dots,x_K)\prod_{k=1}^K h_k(x_k)\,dx_1\cdots dx_K,$$0 as a Gaussian random walk (Chernis et al., 2023).

Dynamic BPS also has a multivariate state-space generalization. In the multivariate macroeconomic formulation, each agent supplies a $$p(y\mid \mathcal H)=\int \alpha(y\mid x_1,\dots,x_K)\prod_{k=1}^K h_k(x_k)\,dx_1\cdots dx_K,$$1-variate density forecast $$p(y\mid \mathcal H)=\int \alpha(y\mid x_1,\dots,x_K)\prod_{k=1}^K h_k(x_k)\,dx_1\cdots dx_K,$$2 for a latent vector, the synthesis density is $$p(y\mid \mathcal H)=\int \alpha(y\mid x_1,\dots,x_K)\prod_{k=1}^K h_k(x_k)\,dx_1\cdots dx_K,$$3, and a structured dynamic latent-factor model yields a DFSUR representation $$p(y\mid \mathcal H)=\int \alpha(y\mid x_1,\dots,x_K)\prod_{k=1}^K h_k(x_k)\,dx_1\cdots dx_K,$$4, $$p(y\mid \mathcal H)=\int \alpha(y\mid x_1,\dots,x_K)\prod_{k=1}^K h_k(x_k)\,dx_1\cdots dx_K,$$5 with random-walk evolution for $$p(y\mid \mathcal H)=\int \alpha(y\mid x_1,\dots,x_K)\prod_{k=1}^K h_k(x_k)\,dx_1\cdots dx_K,$$6 (McAlinn et al., 2017). This retains the basic synthesis logic while allowing dynamic inter-series dependence through $$p(y\mid \mathcal H)=\int \alpha(y\mid x_1,\dots,x_K)\prod_{k=1}^K h_k(x_k)\,dx_1\cdots dx_K,$$7 and through the joint posterior of the synthesis coefficients (McAlinn et al., 2017).

3. Sequential learning and posterior computation

Posterior computation in dynamic BPS is driven by the latent agent states and the time-varying synthesis parameters. In the linear-Gaussian case, the standard posterior target is

$$p(y\mid \mathcal H)=\int \alpha(y\mid x_1,\dots,x_K)\prod_{k=1}^K h_k(x_k)\,dx_1\cdots dx_K,$$8

and McAlinn and West use a two-block Gibbs scheme: conditional on $$p(y\mid \mathcal H)=\int \alpha(y\mid x_1,\dots,x_K)\prod_{k=1}^K h_k(x_k)\,dx_1\cdots dx_K,$$9, the model is a conditionally Gaussian DLM and $\alpha(y\mid x_{1:K})$0 are sampled by forward filtering–backward sampling (FFBS); conditional on $\alpha(y\mid x_{1:K})$1, each $\alpha(y\mid x_{1:K})$2 is sampled from

$\alpha(y\mid x_{1:K})$3

with closed-form Gaussian updates when the agent densities are normal and scale-mixture augmentation when they are Student-$\alpha(y\mid x_{1:K})$4 (McAlinn et al., 2016).

A conjugate filtering form is available for the DLM synthesis. With a normal–inverse–gamma prior on $\alpha(y\mid x_{1:K})$5, forward filtering updates the prior moments $\alpha(y\mid x_{1:K})$6, constructs the one-step predictive distribution, and then applies Kalman-style posterior recursions for $\alpha(y\mid x_{1:K})$7 (Kato et al., 8 Oct 2025). This yields a fully Bayesian sequential updating scheme in which the synthesis coefficients are literally time-varying intercept and weights on the component forecasts (Kato et al., 8 Oct 2025).

Outcome-dependent dynamic BPS requires richer latent structure. In the mixture-based formulation with dynamic biases, covariance parameters, and Dirichlet base weights, posterior inference targets $\alpha(y\mid x_{1:K})$8 using a Gibbs sampler augmented by a discrete indicator selecting which mixture component “fired.” To maintain conjugate parametric priors for the next time point, the Monte Carlo samples are re-projected to a Normal-Inverse-Wishart distribution for $\alpha(y\mid x_{1:K})$9 and a Dirichlet distribution for $y$0 by minimizing Kullback–Leibler divergence, a step described as “variational Bayes” in the paper (1803.01984).

Sequential computation has also been specialized to avoid repeated full MCMC. For linear and Gaussian synthesis, a Rao-Blackwellized particle filter marginalizes out $y$1 and tracks only a particle approximation to $y$2, using the closed-form Student-$y$3 one-step predictive density as the particle weight. When the effective sample size falls below a threshold, the algorithm discards the particle set and invokes an MCMC approximation to the posterior, then restarts the filter with equal-weight particles (Masuda et al., 2023). In the reported U.S. quarterly inflation example, this RBCPF plus timely MCMC intervention maintained synthesized-posterior accuracy at a small fraction of the cost of MCMC-only computation, reported as approximately $y$4 seconds versus $y$5 seconds per time point (Masuda et al., 2023).

The same paper proposes adaptive discount-factor estimation through power-discounted likelihoods. A grid of discount-factor pairs is evaluated by discounted log predictive likelihoods, and softmax or argmax weights over the grid define a second-layer mixture predictive density. In the 2020–2022 inflation burst, fixed-discount DBPS adapted slowly, whereas discount-factor estimation through the Loss Discounting Framework produced wider intervals peaking at the burst and improved log-predictive-density ratios (Masuda et al., 2023).

4. Synthesis-function families and major variants

Dynamic BPS is defined by the synthesis function $y$6, and much of the methodology consists of specifying that function so as to capture calibration structure that static linear pools cannot represent.

Variant	Synthesis function	Distinctive feature
Outcome-dependent pool	Mixture with $y$7 and optional baseline $y$8	Weights depend on forecast outcome
Linear-Gaussian dynamic BPS	$y$9	Dynamic intercepts and weights in a DLM
Tree-based BPS-RT	Regression-tree prior means for $x_{1:K}$0	Nonparametric dependence on weight modifiers
DRQS / FDRQS	Asymmetric Laplace synthesis	Direct quantile targeting
Network BPS	Affine-logit pool $x_{1:K}$1	Dynamic mechanism weights on edges

In outcome-dependent BPS, the synthesis density can be written in discrete-mixture form as

$x_{1:K}$2

which implies

$x_{1:K}$3

The baseline density $x_{1:K}$4 is included to address model-set incompleteness, and the functions $x_{1:K}$5 defining $x_{1:K}$6 can take Gaussian “well,” softmax, “consensus,” or “herding” forms (1803.01984). Because $x_{1:K}$7 depends on the realized value of $x_{1:K}$8, this formulation permits regional calibration in the tails or in other outcome regimes (1803.01984).

Tree-based BPS-RT replaces the static or random-walk prior on synthesis weights by nonparametric functions of user-chosen “weight modifiers” such as ex-post forecast scores, moments of the agent densities, cross-sectional dispersion, financial conditions, uncertainty, or a time trend. The prior means $x_{1:K}$9 and $h_j$0 are learned by regression trees or sums of trees, while local scaling parameters follow horseshoe priors (Chernis et al., 2023). Although the tree structures are fixed through time, the modifiers $h_j$1 vary with $h_j$2, so the implied combination weights vary with time; when an indicator crosses a tree threshold, the prior mean for a dynamic coefficient can jump from one leaf value to another (Chernis et al., 2023).

Dynamic Bayesian regression quantile synthesis (DRQS) changes the synthesis target from conditional means to conditional quantiles by using the asymmetric Laplace distribution in the observation equation,

$h_j$3

so that $h_j$4 is the $h_j$5th conditional quantile of $h_j$6 (Kobayashi et al., 12 Mar 2026). Its multivariate extension, factor DRQS (FDRQS), imposes a time-varying latent factor structure on the synthesis weights to leverage cross-sectional dependencies across multiple time series (Kobayashi et al., 12 Mar 2026).

A further extension adapts BPS to dynamic networks. In that setting, each structural mechanism—such as a stochastic block model, generalized random dot-product graph, Chung–Lu model, or Adamic–Adar score—is treated as an agent returning edge-probability forecasts. The synthesis layer stacks the agent logits into features $h_j$7 and models edge probabilities through

$h_j$8

with $h_j$9 (Papamichalis et al., 18 Jun 2026). This is a dynamic affine-logit pool rather than a Gaussian regression, but it preserves the core BPS architecture of latent agent forecasts plus dynamic synthesis weights (Papamichalis et al., 18 Jun 2026).

5. Empirical domains and observed behavior

The original dynamic BPS time-series study examined quarterly U.S. inflation, short-term interest rates, and unemployment from 1961Q1 to 2014Q4 using four agent DLMs. In the 1990Q1–2014Q4 evaluation period, one-step BPS achieved $x_{1:J}$0, compared with $x_{1:J}$1 for the next best agent, and the horizon-customized four-step model BPS(4) achieved $x_{1:J}$2 compared with $x_{1:J}$3 for direct BPS, with LPDR gains of order $x_{1:J}$4 to $x_{1:J}$5 nats over single agents at horizon $x_{1:J}$6 (McAlinn et al., 2016). The same study reports that BPS and especially BPS(4) dominate Bayesian Model Averaging, linear and log pools, and the DeCo approach, with particularly marked differences during crisis periods (McAlinn et al., 2016).

In the outcome-dependent pooling paper, dynamic BPS is illustrated on Euro–USD daily log-prices over July–December 2016, forecasting five trading days ahead with time-varying autoregressions and DLMs as agents. Using “consensus” weights, posterior trajectories reveal evolving model biases, time-varying cross-model correlations that hit lows around the U.S.-election volatility and recover thereafter, and adaptive base weights that translate into effective outcome-dependent weights in the MCMC draws. In that setting, BPS outperforms each individual model, a static equal-weight pool, classical BMA scoring one day ahead, and an equal-weight pool augmented with baseline density (1803.01984).

Chernis, McAlinn, and collaborators evaluate tree-based dynamic BPS in two macroeconomic applications. For euro-area GDP growth using the ECB Survey of Professional Forecasters at the two-quarter horizon, a single-tree BPS-RT with stochastic volatility and AVG-SCORES weight modifiers achieves approximately $x_{1:J}$7–$x_{1:J}$8 lower CRPS than BPS-RW, with statistical significance by Diebold–Mariano, similar RMSE to benchmarks, well-calibrated PIT histograms, and sparse tree structures with one to two splits capturing crisis versus tranquil periods. For U.S. inflation from FRED-QD indicators, single-tree BPS-RT cuts CRPS by approximately $x_{1:J}$9–$h_k(x_k)$00 relative to BPS-RW at horizon $h_k(x_k)$01, with larger improvements at horizon $h_k(x_k)$02, and adapts quickly to the 2021–2022 inflation spike (Chernis et al., 2023).

Quantile-targeted dynamic BPS has been tested on U.S. inflation-at-risk and global GDP growth-at-risk. In the inflation application, DRQS achieves lower relative cumulative scores especially in the right tail and at the four-quarter horizon, broadening intervals appropriately and avoiding extreme mis-forecasting seen in some agents. In the global GDP application, FDRQS attains the lowest total relative cumulative score at both horizons, outperforms all agents and univariate DRQS, and during the COVID-19 shock in 2020 its RTCS “barely rises” while DRQS and the agent models exhibit large score deteriorations (Kobayashi et al., 12 Mar 2026).

Recent work also carries dynamic BPS into domains beyond univariate macroeconomic density combination. In network forecasting, the synthesis layer returns calibrated edge forecasts and inference on mechanism weights, and the theory proves that a single snapshot identifies and estimates the weights at rate $h_k(x_k)$03, with a sharp distinguishability threshold $h_k(x_k)$04 and a tracking bound that pays only at regime switches (Papamichalis et al., 18 Jun 2026). In portfolio analysis, predictive-synthesis ideas have been extended in two related directions: Bayesian Predictive Decision Synthesis tilts model-specific predictive densities by realized utility scores in sequential portfolio rebalancing (Tallman et al., 2024), and a later study uses BPS with dynamic linear models to combine multiple asset-return prediction models and then construct mean-variance and quantile-based portfolios from the synthesized posterior (Kato et al., 8 Oct 2025).

6. Conceptual relations, misconceptions, and active directions

Dynamic BPS is broader than classical Bayesian Model Averaging. The outcome-dependent pooling paper states explicitly that classical BMA is a special case in which the outcome-dependent synthesis weights reduce to outcome-independent BMA weights and no baseline density is used (1803.01984). By contrast, BPS can incorporate a baseline or “safe-haven” density to address model-set incompleteness, can calibrate models differently across outcome regions, and can represent dependence structures among forecasts rather than treating models as competing, mutually exclusive explanations (1803.01984).

A common misconception is that the factorized term $h_k(x_k)$05 implies that the DM assumes the agents are independent. Masuda and Irie clarify that no assumption of agent-forecast independence is required in the DM’s actual prior; independence appears only in one proof as a convenient artificial prior sharing the same marginal means, and in the graphical interpretation conditional on the reported forecasts (Masuda et al., 2024). Another misconception is that dynamic BPS is synonymous with random-walk Gaussian regression. The literature now includes outcome-dependent pools, regression-tree synthesis, multivariate latent-factor forms, asymmetric-Laplace quantile synthesis, affine-logit network synthesis, and sequential particle-filter implementations (1803.01984, Chernis et al., 2023, Kobayashi et al., 12 Mar 2026, Papamichalis et al., 18 Jun 2026, Masuda et al., 2023).

The current methodological frontier is heterogeneous. In BPS-RT, the reported best empirical performance came from $h_k(x_k)$06 rather than large BART-style ensembles, which the authors interpret as suggesting that combination weights are largely stable with occasional regime shifts; they propose learning $h_k(x_k)$07 automatically or using spike-and-slab priors to control complexity (Chernis et al., 2023). The same paper notes that weight modifiers must be chosen by the user, that multivariate BPS-RT is a natural extension, and that sequential Monte Carlo may be preferable in real-time or very high-frequency settings (Chernis et al., 2023). DRQS and FDRQS push BPS toward direct quantile synthesis and cross-series factor structures (Kobayashi et al., 12 Mar 2026), while the dynamic-network paper adds explicit identification theory and per-switch localization results (Papamichalis et al., 18 Jun 2026). Taken together, these developments indicate that dynamic BPS functions less as a single model class than as a coherent synthesis principle: latent agent forecasts are combined through a Bayesian calibration layer whose structure, dynamics, and inferential machinery can be adapted to the forecasting target at hand.