I am presenting my latest paper in the International Symposium on Forecasting in Montreal, Canada, 2026. The paper theorises different strategies to handle a disruption using the bias-variance decomposition. The abstract is as follows:
This study examines various forecasting strategies in the presence of a disruption, using Hong Kong tourism demand as a case study. The following strategies consist of (1) developing adaptive models, (2) incorporating a disruption into an adaptive model, (3) interpolating the interrupted period of the demand, and (4) controlling the size of the training sets. The bias-variance trade-off provides a theoretical understanding of when a strategy produces more accurate forecasts than other strategies. The study employs single-source-of-error state-space models, namely exponential smoothing and autoregressive integrated moving average, to produce forecasts. This study also assesses their performance relative to the seasonal naive. The findings show that, regardless of the strategies, using all observations is preferable to using post-disruption observations in terms of forecast accuracy. Using mixed-effects regression analysis, imputing the interrupted demand improves the forecast accuracy the most, even though the imputation may introduce errors. If a modeller decides to use only post-disruption data, using a predetermined model structure and backcasting are the best strategies. Based on these findings, a practical decision tree for modelling is proposed.
The framework is proposed as follows:
The working paper is available on this link. Looking forward to discussing more!
Leave a Reply