BSTS (Bayesian Structural Time Series)

BSTS (Bayesian Structural Time Series) is a probabilistic time series forecasting model that combines structural decomposition with Bayesian inference.

It is widely used for: - Forecasting with uncertainty - Causal impact analysis - Handling missing data - Complex time series patterns

Model Type

  • Category: Bayesian Forecasting Model
  • Framework: State Space Model
  • Inference Method: MCMC (Markov Chain Monte Carlo)
  • Paradigm: Probabilistic / Stochastic

BSTS decomposes a time series into interpretable components and models them probabilistically.

Instead of producing a single forecast, it generates a distribution of possible future outcomes using Bayesian simulation.

Key Components

1. Level (Local Level)

Represents the baseline value of the series.

2. Trend

Captures upward or downward movement over time.

3. Seasonality

Models repeating patterns (daily, weekly, yearly).

4. Regression Components (Optional)

External predictors (covariates) can be added for better accuracy.

How It Works

  1. The time series is represented as a state-space model
  2. Priors are defined for components (level, trend, seasonality)
  3. MCMC sampling is used to simulate posterior distributions
  4. Thousands of possible future paths are generated
  5. Final forecast is derived from aggregated simulations