Dynamic Bayesian Model Averaging
Also known as: DMA, dynamic model averaging, time-varying BMA, online Bayesian model averaging
Dynamic Bayesian Model Averaging (DMA) extends standard Bayesian model averaging to settings where the best predictive model may change over time. It maintains a probability distribution over a set of competing models and updates that distribution sequentially as new observations arrive, allowing model weights to evolve rather than remaining fixed across the entire sample.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use DMA when you have sequential data and suspect the predictive relationship changes over time — economic forecasting with regime shifts, sensor streams where calibration drifts, or any setting where no single model can be assumed stable across the whole sample. It is especially useful when the candidate model set spans different covariate combinations and you want data-driven variable selection that is allowed to evolve. Do not use it as a drop-in for standard BMA on cross-sectional data: the forgetting mechanism is designed for ordered sequences. Avoid it when the model set is very large (exponential in the number of predictors) without a computationally efficient approximation, and when interpretability of a single stable model is paramount.
Strengths & limitations
- Allows model weights to change over time, capturing structural breaks and regime shifts.
- Combines model uncertainty with parameter uncertainty within a coherent Bayesian framework.
- Produces well-calibrated probabilistic forecasts by averaging over the full model set.
- Computationally tractable via the forgetting-factor recursion, even with moderate model sets.
- Provides time-varying variable inclusion probabilities, revealing which predictors matter at each period.
- The forgetting factor alpha must be set or tuned; it is not estimated from the data by default.
- The model set must be specified in advance; models omitted from consideration can never be selected.
- Computational cost grows with the number of candidate models, becoming impractical for very large sets without approximations.
- Interpretability is reduced compared to a single model, since the active specification shifts over time.
Frequently asked
How does DMA differ from standard Bayesian model averaging?
Standard BMA computes fixed model weights using all available data and does not allow those weights to change over time. DMA applies a forgetting factor that progressively discounts older evidence, enabling the weight on each model to track changes in the data-generating process. For stationary data with no structural change, both approaches converge to similar answers.
How do I choose the forgetting factor alpha?
A common practice is to evaluate predictive log-scores on a hold-out or rolling validation window for a grid of alpha values (e.g. 0.90, 0.95, 0.99, 1.00) and select the value that maximises average predictive accuracy. Alpha near 1 is appropriate for slow-changing or stationary processes; alpha closer to 0.9 allows faster adaptation but introduces more variance.
What is the model set and how large can it be?
The model set is the collection of all candidate specifications the algorithm considers. With p potential predictors, the full set has 2^p members, which becomes computationally prohibitive beyond roughly 20 predictors. Approximations such as the dynamic occam's window or stochastic search over model subspaces are used for larger problems.
Does DMA handle parameter change within each model?
DMA as originally formulated averages over a fixed set of models; parameters within each model may be static or time-varying depending on how each model is specified. Combining DMA with state-space or Kalman-filter-based models allows both model weights and parameters to evolve simultaneously.
Is DMA the same as dynamic model selection?
They are related but distinct. Dynamic model selection (DMS) always assigns full weight to the single highest-probability model at each step, while DMA maintains a mixture over all models. DMA generally produces better-calibrated probabilistic forecasts; DMS is more interpretable but can be unstable when two models have similar probabilities.
Sources
- Raftery, A. E., Karny, M., & Ettler, P. (2010). Online prediction under model uncertainty via dynamic model averaging: Application to a cold rolling mill. Technometrics, 52(1), 52-66. DOI: 10.1198/TECH.2009.08104 ↗
- Hoeting, J. A., Madigan, D., Raftery, A. E., & Volinsky, C. T. (1999). Bayesian model averaging: A tutorial. Statistical Science, 14(4), 382-401. link ↗
How to cite this page
ScholarGate. (2026, June 3). Dynamic Bayesian Model Averaging. ScholarGate. https://scholargate.app/en/bayesian/dynamic-bayesian-model-averaging
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Bayesian Model AveragingBayesian↔ compare
- Dynamic Bayesian InferenceBayesian↔ compare
- Dynamic Bayesian NetworkBayesian↔ compare
- Dynamic Variational InferenceBayesian↔ compare
- Kalman FilterBayesian↔ compare
- Sequential Monte CarloBayesian↔ compare