Bayesian structural time series (BSTS) is a statistical framework for modeling and forecasting time series data. It integrates state-space models with Bayesian inference, allowing for the decomposition of a series into trend, seasonal, and regression components while providing full posterior distributions for all parameters. This approach is particularly valued for its ability to handle missing data, incorporate prior information, and quantify uncertainty in predictions and causal effect estimates.
The method was popularized in the mid-2010s through academic work and open-source implementations, notably the CausalImpact package developed by Google. BSTS has become a standard tool in fields such as econometrics, epidemiology, and marketing analytics, where understanding the impact of interventions or events on a time series is critical. Its flexibility and probabilistic nature distinguish it from classical time series methods like ARIMA, which often require stationarity and treat parameters as fixed.
Core Components
BSTS models are built on a state-space representation, where an observed time series \( y_t \) is expressed as a function of a latent state vector \( \alpha_t \). The observation equation is typically \( y_t = Z_t^T \alpha_t + \varepsilon_t \), and the state evolves according to \( \alpha_{t+1} = T_t \alpha_t + R_t \eta_t \), with error terms \( \varepsilon_t \) and \( \eta_t \) assumed to be Gaussian. The state vector can include multiple components: a local linear trend, a seasonal component (often modeled with Fourier terms or dummy variables), and regression coefficients for external predictors.
A key feature is the use of spike-and-slab priors on regression coefficients. This prior distribution allows the model to automatically perform variable selection, shrinking irrelevant coefficients to zero while retaining important predictors. This is especially useful when the number of potential covariates is large relative to the sample size, a common scenario in observational studies.
Bayesian Inference and Computation
Inference in BSTS is conducted through Markov chain Monte Carlo (MCMC) methods, typically using Gibbs sampling. The state-space structure allows for efficient sampling of latent states via the Kalman filter and smoother, while the regression coefficients and variance parameters are updated conditionally. The Bayesian approach yields a posterior distribution for every quantity, enabling credible intervals for forecasts and effect estimates.
One practical advantage is the natural handling of missing observations. In a state-space framework, missing values are treated as latent variables and are imputed during the sampling process, avoiding the need for separate imputation steps. This is particularly beneficial for real-world datasets with gaps due to reporting delays or data collection issues.
Causal Impact Analysis
A prominent application of BSTS is causal impact estimation, as implemented in the CausalImpact R package. The method constructs a synthetic control for a treated unit by fitting a BSTS model to a set of control time series during a pre-intervention period. The model learns the relationship between the treated series and controls, then predicts the counterfactual outcome for the post-intervention period. The difference between observed and predicted values provides an estimate of the causal effect, complete with uncertainty intervals.
This approach has been widely used in evaluating the effect of marketing campaigns, policy changes, or public health interventions. For example, researchers have applied it to measure the impact of advertising spend on sales or the effect of a new law on traffic accidents. The method's strength lies in its ability to account for trends and seasonality without requiring a randomized experiment.
Software Implementations
Several software packages implement BSTS. The original CausalImpact package, released by Google in 2014, remains popular in R. The bsts R package, also from Google, provides a more general framework for building custom state-space models. In Python, the pycausalimpact library offers a port of CausalImpact, and the statsmodels library includes some state-space functionality, though with less emphasis on Bayesian variable selection.
These tools have lowered the barrier to entry, allowing practitioners without deep expertise in Bayesian statistics to apply BSTS. However, users must still make modeling choices, such as the number of seasonal components or the prior distribution for regression coefficients, which can influence results.
Limitations and Considerations
BSTS is computationally intensive compared to classical methods, especially for long series or many covariates, due to the MCMC sampling. Convergence diagnostics are necessary to ensure reliable posterior estimates. The method also assumes linearity and Gaussian errors, which may not hold for count data or highly skewed distributions, though extensions exist for Poisson and other families.
Another limitation is the reliance on the choice of control series in causal applications. If controls are poor predictors or are themselves affected by the intervention, the counterfactual estimates can be biased. Despite these challenges, BSTS remains a robust and flexible tool, bridging the gap between traditional econometrics and modern machine learning approaches.
Related Concepts
BSTS is related to broader fields of Machine learning and Artificial intelligence, particularly in forecasting applications. It shares conceptual ground with Sequence-to-Sequence (Seq2Seq) models and Encoder-Decoder Architecture architectures, which also handle temporal dependencies, though BSTS emphasizes interpretability and uncertainty quantification. The method also connects to Data Augmentation techniques, as missing data imputation is a core feature. In the context of Generative AI, BSTS can be used to generate realistic counterfactual scenarios, though it is not a neural-network-based approach.