On forecast stability

Abstract

Forecasts are typically produced in a business context on a regular basis to make downstream decisions. Here, forecasts should not only be as accurate as possible, but also should not change arbitrarily, and be stable in some sense. In this paper, we explore two types of forecast stability that we call vertical stability (for forecasts from different origins for the same target) and horizontal stability (for forecasts from the same origin for different targets). Existing works in the literature are only applicable to certain base models and can only stabilise forecasts vertically. We propose a simple linear-interpolation-based approach to stabilise the forecasts provided by any base model, both vertically and horizontally. Our method makes the trade-off between stability and accuracy explicit, producing forecasts at any point in the spectrum of this trade-off. We used N-BEATS, pooled regression, LightGBM, ETS, and ARIMA as base models in our evaluation across different error and stability measures on four publicly available datasets. On some datasets, the proposed framework achieved forecasts that were both more accurate and stable than the base forecasts. On the others, we achieved forecasts that were slightly less accurate but much more stable.