Forecasting seasonal changepoint models utilizing changepoint uncertainties

Forecasting in the presence of changepoints is a challenging task. However, if changepoints occur regularly across seasons, we want to utilize that information in our predictions for the next season. In this paper, we develop a framework for forecasting seasonal data that contain within-season changepoints. In our forecasts, we account for uncertainty in the changepoint locations both within and across seasons via a weighting approach across changepoint uncertainties. The framework is very flexible across different changepoint model assumptions and approaches to identifying changepoint confidence sets. We demonstrate the improvement in forecasting performance of our framework for atmospheric pressure with a multivariate changepoint approach, and for air pollution ( PM 2.5 ) with a univariate changepoint approach.