GoodsForecast second-place solution in M5 Uncertainty track: Combining heterogeneous models for a quantile estimation task

Abstract

GoodsForecast’s predictions obtained a second place in the M5 Uncertainty Competition. This result was achieved using an empirical method that estimates the probability distribution of time series by combining the forecasts of three methods. The first method considers past observations, estimating their histogram, the second method is used to estimate trends through a singular spectrum analysis, and the last one captures external factors with the help of a machine learning model. Forecasts made by these methods at different levels of aggregation were judgmentally combined to account for the particularities of the M5 data set. Quantiles, estimated by the combination of algorithms, were modified with rounding heuristics that resulted in the improvement of the competition target loss metric. This paper is devoted to a detailed description of the developed algorithmic scheme for evaluating the probability distribution, the experiments, and some insights gained by the authors during the problem- solving process.