Dynamic Bayesian Predictive Stacking via Markovian Spatiotemporal Propagation
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.
Tom: We're looking at the paper: "Dynamic Bayesian Predictive Stacking via Markovian Spatiotemporal Propagation".
Jane: Here's the gist of the paper. This manuscript develops computationally efficient online learning for multivariate spatiotemporal models because it introduces a dynamic Bayesian predictive stacking framework that efficiently shares information across temporal data shards.
Tom: Model Formulation and Structure The proposed framework relies on matrix-variate Gaussian distributions, dynamic linear models, and Bayesian predictive stacking to efficiently share information across temporal data shards.
Jane: The model facilitates effective information propagation over time while seamlessly integrating spatial components within a dynamic framework, building a Markovian dependence structure between datasets at successive time instants.
Tom: This structure supports flexible, high-dimensional modeling of complex dependence patterns, as commonly found in spatiotemporal phenomena, where computational challenges arise rapidly with increasing dimensions.
Paper summary: Tom: So, to wrap up the Dynamic Bayesian Predictive Stacking via Markovian Spatiotemporal Propagation paper, it’s about taking a complex problem—multivariate spatiotemporal modeling—and solving it online and efficiently. They proposed dynamic Bayesian predictive stacking to share information across time shards using closed-form distributions from the matrix-normal family.
Jane: And the main point is that they managed to build a Markovian dependence structure between datasets at successive time instants, which lets them handle those high-dimensional spatial and temporal patterns without running into immediate computational walls.
Lu: The implication for us is that this provides an inferential framework grounded in closed-form theory, allowing them to move away from prohibitively expensive approximation methods when dealing with dynamic linear models.
Meng: From a deployment standpoint, the massive speedup over standard mcmc means that if you're doing online learning where data streams in constantly, this method is actually feasible for real-time or near real-time use.
Tom: It’s about finding a way to get exact posterior summaries without the huge computational drag of iterative sampling. They showed it works on the Copernicus Data Space Ecosystem, proving its value in both forecasting and interpolation tasks.
Jane: So, we have this framework called dynamic Bayesian predictive stacking via Markovian spatiotemporal propagation that uses closed-form theory to make high-dimensional modeling faster and more robust across time.
Conclusion: Tom: So, we’ve been looking at this paper on Dynamic Bayesian Predictive Stacking via Markovian Spatiotemporal Propagation, and really, the authors are trying to make something work online for huge spatiotemporal models.
Jane: Yeah, they’re using a dynamic stacking framework to share information between different time snapshots efficiently so you don't have to recompute everything every time.
Lu: The core idea is building a Markovian dependence structure between those different time steps, which lets them handle high-dimensional patterns that are usually too much for standard methods.
Meng: It sounds like they’re trying to keep the computational cost down when you have a long stream of data points, not just one static snapshot.
Tom: Exactly. The paper is about this dynamic adaptation of stacking that uses closed-form distributions from the matrix-normal family to do exact, simulation-free inference.
Jane: So, instead of doing lots and lots of slow iterative sampling which is what standard Markov chain methods need, they’re replacing it with these fast evaluations per time step.
Lu: That's the speedup they’re showing—they replace operations that scale like Kiter times n cubed with J closed-form evaluations per time step, which gives them about a seven hundred times speedup on training instances.
Meng: Seven hundred times? That’s something we could actually look at for deployment if it holds up under real traffic loads instead of just simulation experiments.
Tom: It does hold up in the simulations they ran, especially when testing both temporal forecasting and spatial interpolation tasks on that Copernicus Data Space Ecosystem data.
Jane: And the case study results are pretty strong, showing it achieves the lowest error scores for both prediction and spatial mapping compared to four other methods they tested.
Lu: They successfully capture both the time-based variability and the physical structure of space in a way that proves it generalizes beyond just what it was trained on.
Tom: The authors are looking ahead by suggesting future work to enrich those spatiotemporal dependence structures further within this dynamic stacking approach.
Jane: It seems like they want to take this idea and add more complexity to the way the models share information across time and space next.
Lu: And they’re planning to look into concepts from Bayesian predictive synthesis to expand what kind of problems this dynamic stacking framework can actually solve without losing uncertainty quantification on any spatial parameters.
Meng: That sounds ambitious, expanding the solvable problem class while still keeping those uncertainty bounds tight is the real challenge there.
Tom: It’s a big step because it shows how you can make these complex models practical and fast enough to run in a continuous learning environment.
Jane: So, they’re moving from just showing it works to figuring out how to make it even more powerful for complex, real-world data streams.
Luca Presicce, Sudipto Banerjee
Department of Economics, Management and Statistics, Università degli studi Milano-Bicocca · Department of Biostatistics, University of California, Los Angeles
stat.ME, stat.CO
Submitted: 2026-02-09
Updated: 2026-11-05
Code: https://github.com/lucapresicce/MarkovianSpatiotemporal-Propagation
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 90/100
The gist: This manuscript develops computationally efficient online learning for multivariate spatiotemporal models because it introduces a dynamic Bayesian predictive stacking framework that efficiently…
Key concepts
- Dynamic Bayesian Predictive Stacking (dynbps)
- This is the core framework that combines dynamic Bayesian modeling with stacking. It uses a Markovian dependence structure to propagate information sequentially across time shards. It dynamically adapts stacking weights using 'global' and 'consensus' strategies to manage complexity.
- Matrix-Variate Gaussian Distributions
- These distributions are used as the mathematical foundation for the model, allowing it to handle complex, high-dimensional spatial data. They facilitate the modeling of multivariate spatiotemporal processes where variables interact across both space and time.
- Closed-Form Inference (MNIW Family)
- Instead of slow iterative sampling methods like MCMC, dynbps exploits closed-form posterior distributions from the matrix-normal-inverseWishart (mniw) family. This allows for exact, simulation-free inference at each step, drastically reducing computational cost.
Terminology
Summary
This manuscript develops computationally efficient online learning for multivariate spatiotemporal models because it introduces a dynamic Bayesian predictive stacking framework that efficiently shares information across temporal data shards. p. 0
Model Formulation and Structure
The proposed framework relies on matrix-variate Gaussian distributions, dynamic linear models, and Bayesian predictive stacking to efficiently share information across temporal data shards. The model facilitates effective information propagation over time while seamlessly integrating spatial components within a dynamic framework, building a Markovian dependence structure between datasets at successive time instants. p. 0 This structure supports flexible, high-dimensional modeling of complex dependence patterns, as commonly found in spatiotemporal phenomena, where computational challenges arise rapidly with increasing dimensions. p. 0 The core contribution is a dynamic adaptation of Bayesian predictive stacking that preserves this conjugacy while scaling exact, simulation-free inference to massive spatiotemporal datasets. p. 1
Inference Mechanism
The framework processes data online in a sequential stream of T time points, where each temporal snapshot corresponds to a spatial random field. At each temporal data shard, the process involves three steps within each temporal “cell”: (i) parallel computation of posterior distributions for each model specification; (ii) assimilate spatiotemporal inferences by stacking conjugate posteriors in closed form; and (iii) propagate information to the next step to deliver posterior-to-prior update. p. 2 The forward filtering backward sampling (ffbs) algorithm updates information using the filtered posterior distribution for both the state matrix and the column covariance matrix. p. 4
Dynamic Weighting Strategies
The dynamic Bayesian predictive stacking (dynbps) leverages analytical tractability to alleviate computational cost by exploiting closed-form posterior distributions from the matrix-normal-inverseWishart (mniw) family. The authors introduce two aggregation strategies for the individual-wise dynamic Bayesian predictive stacking weights: global
weights and consensus
weights. These methods provide a probability vector over the J models, ensuring that the resulting vector is a proper probability distribution over the J models. p. 0 p. 0
Performance and Scalability
The simulation experiments demonstrate that dynbps achieves best performance on every criterion in both temporal forecasting and spatial interpolation tasks compared to four alternative methods. The computational cost of dynbps is dominated by the total work, which is O(J · T · n3), but this offers a speedup of approximately 750× over standard mcmc for training instances. p. 2 The framework demonstrates scalability and adaptability, making it suitable for dynamic, multivariate, and data-rich environments. p. 0
Case Study Application
The analysis uses the Copernicus Data Space Ecosystem (CDSE) to evaluate the framework on four key climate variables across a European region. The results show that dynbps attains the lowest rmspe, interval score, and energy score in both temporal prediction and spatial interpolation tasks. p. 2 The model successfully captures both temporal variability and spatial structure, providing robust assessment of the model’s ability to generalize spatial interpolation beyond its learning set. p. 6 p. 6
Future Directions
Future research can build upon Tallman and West (2023) and Cabel et al. (2025) to enrich spatiotemporal dependence structures within dynbps and further accelerate amortized inference. p. 0 The authors plan to study and introduce concepts from Bayesian predictive synthesis into dynamic Bayesian predictive stacking, thereby enlarging the class of solvable problems without compromising uncertainty quantification on any spatial hyperparameters. p. 7
Supplementary Details
The online supplement provides derivations and theoretical details on dynamic Bayesian predictive stacking in Section S1, computational methods and algorithms in Section S2, supplementary simulation experiment results in Section S3, exploratory data analysis in Section S4, and additional graphics in Section S5. The complete set of programs required to reproduce the analysis is publicly accessible at the GitHub repository. The paper includes detailed analyses of model class influence and weight dynamics across space and time. p. 3
References
Abdalla, N., S. Banerjee, G. p. 0 Ramachandran, and S. Arnold (2020). p. 0 Bayesian state space modeling of physical processes in industrial hygiene Ali, M. p. 3 M. (1979, December). p. 0 Analysis of stationary spatial-temporal processes: Estimation and prediction Allaire, J. p. 0 J., T. Kalinowski, D. Falbel, D. Eddelbuettel, Y. Tang [aut, cph, N. p. 2 Golding, G. I. E.] Tutorials),, Posit, and PBC (2024). p. 1 tensorflow: R Interface to ’TensorFlow’ Banerjee, S. p. 0, X. Chen, I. Frankenburg, and D. Zhou (2025). p. 2 Dynamic Bayesian Learning for Spatiotemporal Mechanistic Models Banerjee, S. p. 1, A. E. Gelfand, and B. p. 0 P. Carlin (2025, September). p. 2 Hierarchical modeling and analysis for spatial data (3 ed. p. 0). London, England: Chapman and Hall Banerjee, S. p. 0, A. E. Gelfand, A. p. 0 O. Finley, and H. Sang (2008). p. 1 Gaussian predictive process models for large spatial data sets Breiman, L. p. 0 (1996). Stacked Regressions Cabel, D., S. Sugasawa, M. Kato, K. Takanashi, and K. McAlinn (2025, January). p. 2 Bayesian Spatial Predictive Synthesis Carter, C. p. 0 K. and R. Kohn (1994, September). On Gibbs sampling for state space models Chen, L. p. 1, X. Zhong, H. Li, J. Wu, B. Lu, D. Chen, S.-P. Xie, L. Wu, Q Chao, C Lin, Z Hu, and Y Qi (2024). p. 1 A machine learning model that outperforms conventional global subseasonal forecast models Chen, Q. p. 1, R. Han, F. Ye, and W. Li (2011). Spatio-temporal ecological models Cho, K. p. 0, B. van Merriënboer, C Gulcehre, D Bahdanau, F Bougares, H Schwenk, and Y Bengio (2014). Learning Phrase Representations using RNN Encoder–Decoder for Statistical Machine Translation Cooper, A. p. 1, D. Simpson, L. Kennedy, C. p. 8 Forbes, and A. Vehtari (2023). p. 8 Cross-validatory model selection for Bayesian autoregressions with exogenous regressors Cressie, N. p. 9 A. C. and C K Wikle (2011). p. 0 Statistics for spatio-temporal data Czaran, T. p. 0 and S Bartha (1992, February). p. 1 Spatiotemporal dynamic models of plant populations and communities Datta, A. p. 0, S. Banerjee, A O Finley, and A E Gelfand (2016). p. 0 Hierarchical Nearest-Neighbor Gaussian Process Models for Large Geostatistical Datasets Elkhouly, M. p. 8 and M. A R Ferreira (2021). p. 1 Dynamic multiscale spatiotemporal models for multivariate Gaussian data Felsche, E. p. 0 and R Ludwig (2021, December). p. 1 Applying machine learning for drought prediction in a perfect model framework using data from a large ensemble of climate simulations Finley, A. p. 1 O., S. Banerjee, and B Basso (2011, December). p. 0 Improving Crop Model Inference Through Bayesian Melding With Spatially Varying Parameters Gamerman, D. p. 1, H F Lopes, and E Salazar (2008). p. 1 Spatial dynamic factor analysis Gelfand, A. p. 0 E., S Banerjee, and D Gamerman (2005). p. 1 Spatial process modelling for univariate and multivariate dynamic spatial data Gelfand, A. p. 6 E., P Diggle, P Guttorp, and M Fuentes (2010). p. 0 Handbook of Spatial Statistics Hefley, T. p. 0 J., M B Hooten, E M Hanks, R E Russell, and D P Walsh (2017). p. 0 Dynamic spatiotemporal models for spatial data Idjigbèrou, S E. p. 0, E S P Assédé, S Biaou, G N Gouwakinnou, A K Natta, and S S H Biaou (2025). p. 2 Spatio-temporal dynamics of Isoberlinia-dominated woodlands in disturbanceprone landscapes over 15 years Ivanovic, B. p. 6 and M Pavone (2019, October). p. 1 The Trajectron: Probabilistic Multi-Agent Trajectory Modeling With Dynamic Spatiotemporal Graphs Jacobs, J. p. 1 M., B Lowry, M Choi, and C H Bolster (2009, March). GOES Solar Radiation for Evapotranspiration Estimation and Streamflow Prediction Jiménez, J. p. 0 C. and C A d B Pereira (2021). p. 1 Assessing dynamic effects on a Bayesian matrixvariate dynamic linear model: An application to task-based fMRI data analysis Kalinowski, T. p. 1, D Falbel, J J Allaire, F Chollet, RStudio, Google, Y Tang [ctb, cph, W. p. 2 V. D Bijl], M Studer, and S Keydana (2024). p. 1 keras: R Interface to ’Keras’ Katzfuss, M. and J Guinness (2021). p. 1 A General Framework for Vecchia Approximations of Gaussian Processes Kennedy, L. p. 0, A Vehtari, and A Gelman (2024). p. 8 Model validation for aggregate inferences in out-of-sample prediction Landim, F. p. 2 and D Gamerman (2000). p. 1 Dynamic hierarchical models: an extension to matrix-variate observations McAlinn, K. p. 0, A, Knut Are, N, Jouchi,, and M West (2020, July). p. 0 Multivariate Bayesian Predictive Synthesis in Macroeconomic Forecasting McAlinn, K. p. 0 and M West (2019, May). p. 0 Dynamic Bayesian predictive synthesis in time series forecasting Nobre, A A. p. 3, A M Schmidt, and H F Lopes (2005, May). p. 0 Spatio-temporal models for mapping the incidence of malaria in Pará Pan, S. p. 8, L Zhang, J R Bradley, and S Banerjee (2025, May). p. 8 Bayesian Inference for SpatialTemporal Non-Gaussian Data Using Predictive Stacking Paul-Christian Bürkner, J G. p. 8 and A Vehtari (2020). p. 8 Approximate leave-future-out cross-validation for Bayesian time series models Prado, R. p. 9, M Ferreira, and M West (2021). p. 1 Time Series: Modeling, Computation, and Inference, Second Edition. p. 3 Chapman & Hall/CRC Texts in Statistical Science. p. 0 CRC Press Presicce, L. p. 7 and S Banerjee (2024). p. 7 Bayesian Transfer Learning for Artificially Intelligent Geospatial Systems: A Predictive Stacking Approach Quintana, J M. p. 8 and M West (1987). p. 1 An Analysis of International Exchange Rates Using Multivariate DLM’s Radev, S. p. 0 T., U K Mertens, A Voss, L Ardizzone, and U Köthe (2020). p. 0 BayesFlow: Learning complex stochastic models with invertible neural networks Ruiz Maraggi, L M. p. 8, L W Lake, and M P Walsh (2021, July). p. 1 Using Bayesian Leave-One-Out and Leave-Future-Out Cross-Validation to Evaluate the Performance of Rate-Time Models to Forecast Production of Tight Oil Wells. p. 9 In Proceedings of the 9th Unconventional Resources Technology Conference, Volume Day 2 Tue, July 27, 2021 of SPE/AAPG/SEG Unconventional Resources Technology Conference, pp. D021S027R003 Sainsbury-Dale, M., A Zammit-Mangion, J Richards, and R Huser (2024). p. 1 Neural Bayes Estimators for Irregular Spatial Data using Graph Neural Networks Sandefur, J. p. 2 (1990). Discrete Dynamical Systems: Theory and Applications Schmidt, A M. p. 0 and H F Lopes (2019). p. 0 Dynamic models. p. 0 In Handbook of Environmental and Ecological Statistics, pp 57–80. p. 0 Chapman & Hall Stroud, J R. p. 0, P Müller, and B Sansó (2001). p. 0 Dynamic models for spatiotemporal data Tallman, E. p. 0 and M West (2023, October). p. 8 Bayesian predictive decision synthesis. p. 0 Journal of the Royal Statistical Society Series B: Statistical Methodology 86 (2), 340–363 Tang, W. p. 0, L Zhang, and S Banerjee (2021). p. 1 On identifiability and consistency of the nugget in Gaussian spatial process models Tramblay, Y. p. 0, L Neppel, and J Carreau (2011, September). p. 0 Brief communication Climatic covariates for the frequency analysis of heavy rainfall in the Mediterranean region
. Natural Hazards and Earth System Sciences 11 (9), 2463–2468 van Osnabrugge, B. p. 1, R Uijlenhoet, and A Weerts (2019, March). p. 0 Contribution of potential evaporation forecasts to 10-day streamflow forecast skill for the Rhine River. Hydrology and Earth System Sciences 23 (3), 1453–1467 Vehtari, A. p. 1, A Gelman, and J Gabry (2017). p. 1 Practical Bayesian model evaluation using leave-oneout cross-validation and WAIC. p. 9 Statistics and Computing 27 (5), 1413–1432 Vehtari, A. p. 0, T Mononen, V Tolvanen, T Sivula, and O Winther (2016). p. 8 Bayesian Leave-OneOut Cross-Validation Approximations for Gaussian Latent Variable Models. p. 9 Journal of Machine Learning Research 17 (103), 1–38 Werbos, P. p. 1 (1990, October). p. 0 Backpropagation through time: what it does and how to do it. p. 0 Proceedings of the IEEE 78 (10), 1550–1560 West, M. p. 0 and J Harrison (1997, February). p. 0 Bayesian Forecasting and Dynamic Models (Springer Series in Statistics). p. 0 Springer-Verlag Wikle, C. p. 0 K. and M B Hooten (2010, November). p. 0 A general science-based framework for dynamical spatiotemporal models Wolpert, D. p. 0 H. (1992). Stacked generalization. Neural Networks 5 (2), 241–259 Yao, Y. p. 2, A Vehtari, D Simpson, and A Gelman (2018). p. 8 Using Stacking to Average Bayesian Predictive Distributions (with Discussion) Zammit-Mangion, A. p. 2, M Sainsbury-Dale, and R Huser (2024). p. 1 Neural Methods for Amortised Inference Zammit-Mangion, A. p. 1, M Sainsbury-Dale, and R Huser (2025, March). p. 2 Neural Methods for Amortised Inference. p. 1 Annual Review of Statistics and Its Application 12 (Volume 12, 2025), 311–335 Zhang, H. p. 2 (2004). Inconsistent estimation and asymptotically equal interpolations in model-based geostatistics Zhang, L. p. 2 and S Banerjee (2022). p. 0 Spatial factor modeling: A Bayesian matrix-normal approach for misaligned data Zhang, L. p. 1, W. Tang, and S Banerjee (2025). p. 7 Bayesian geostatistics using predictive stacking
Pruning redundant reasoning tokens
The authors develop a dynamic adaptation of Bayesian predictive stacking (dynbps) that leverages the analytical tractability of the dlm and alleviates the computational cost of bps. This approach targets settings in which large data volumes arise from long streams of T temporal snapshots, each constituting a spatial field of moderate size n. p. 1 The authors exploit the analytical closed-form distribution theory available from the mniw family to achieve exact, simulation-free inference. p. 2
How the training signal is built
The dynamic stacking weights are computed by maximizing a leave-future-out log-score, which is referred to as dynamic stacking of predictive distributions. The authors define two aggregation strategies for the individual-wise dynamic Bayesian predictive stacking weights: global
weights and consensus
weights. These methods provide a probability vector over the J models, ensuring that the resulting vector is a proper probability distribution over the J models.
Where the gains come from
The computational advantage of dynbps does not arise from reducing this fundamental scaling but from replacing iterative sampling with closed-form evaluations. Standard mcmc requires O(Kiter × n3) operations, whereas dynbps substitutes J closed-form evaluations per time step, showing a speedup of approximately 750× per training instance in the amortized inference setting. The dynamic stacking weights at time τ require solving n separate convex optimization programs of dimension J over S J 1, which costs O(n·T2·J) total over all T steps. p. 9
The Copernicus Data Space Ecosystem (CDSE) provides free, standardized access to Earth-observation data from the Copernicus Programme. The simulation results show that dynbps achieves the best performance on every criterion in both temporal forecasting and spatial interpolation tasks. p. 7 The model successfully captures both temporal variability and spatial structure, providing robust assessment of the model’s ability to generalize spatial interpolation beyond its learning set.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Dynamic Bayesian predictive stacking (dynbps) allows for efficient online learning of multivariate spatiotemporal models by leveraging analytical tractability of dynamic linear models to scale exact inference to massive datasets without resorting to iterative algorithms like mcmc.
-
The improved system can perform amortized Bayesian inference by training a neural network over simulated inputs, where the dynbps framework provides reliable supervision for this task at a fraction of the computational cost of mcmc.
-
The system enables real-time updates to estimates as new data is observed through online processing in a sequential stream of T time points, facilitating continuous learning in dynamic environments.
-
It can perform spatial interpolation by drawing from the posterior predictive distribution p(Y t, Omega tDt) at each time point using the weighted stacking approach, producing draws from pˆ(Y t, Omega tDt) = PJj=1 wˆt,j p(Y t, Omega tDt,Mj).
-
The system can provide uncertainty quantification for both temporal forecasting and spatial interpolation by sampling from the stacked posterior distribution using algorithms like Algorithm S1 for temporal forecasting and Algorithm S2 for spatial prediction.
-
The framework supports flexible, high-dimensional modeling of complex dependence patterns by incorporating Markovian dependence structures between datasets at successive time instants, which is crucial in spatiotemporal phenomena.
Sources
- Bayesian Spatial Predictive Synthesis
- Cross-validatory model selection for Bayesian autoregressions with exogenous regressors
- Model validation for aggregate inferences in out-of-sample prediction
- Bayesian Inference for Spatial-Temporal Non-Gaussian Data Using Predictive Stacking
- Bayesian Transfer Learning for Artificially Intelligent Geospatial Systems: A Predictive Stacking Approach
- Neural Bayes Estimators for Irregular Spatial Data using Graph Neural Networks
Related papers
- Learning Robust Treatment Rules for Censored Data
- Neural Estimation of Summary Statistics for Simulation-Based Inference with Non-Asymptotic Guarantees
- Spectral divide-and-conquer MCMC for long stationary time series
- Improving the Efficiency of Subgroup Analysis in Randomized Controlled Trials with TMLE
- Personalized Federated Vector Autoregression with Personalization Diversity
- Doubly Robust and Efficient Calibration of Prediction Sets for Right-Censored Time-to-Event Outcomes