Evaluating the Robustness of Time Series Forecast Models Under Disruptions

In uncertain, hyper-connected and fast-changing environments, supply chains are becoming more prone to disruptions and unexpected events which affect their stability and ultimately the future consumer demand. Traditional time series forecasting approaches fail to take into consideration environmental uncertainties related to the volatile nature of the supply chains. This paper provides an evaluation of the robustness of time series forecast models under disruptions and future uncertainties. By introducing disruption scenarios with varying intensities and durations, the study evaluates the robustness of forecast models using a range of statistical and machine learning models. This study was evaluated on M3 monthly data to test the performance of forecast models and selection strategies using various accuracy metrics. The results underscore the need for a new robust forecast model selection approach to find a trade-of between accuracy and robustness to future uncertainties.