Resampling strategies for imbalanced time series

Abstract

Time series forecasting is a challenging task, wherethe non-stationary characteristics of the data portrays a hardsetting for predictive tasks. A common issue is the imbalanceddistribution of the target variable, where some intervals are veryimportant to the user but severely underrepresented. Standardregression tools focus on the average behaviour of the data.However, the objective is the opposite in many forecasting tasksinvolving time series: predicting rare values. A common solutionto forecasting tasks with imbalanced data is the use of resamplingstrategies, which operate on the learning data by changing its dis-tribution in favor of a given bias. The objective of this paper is toprovide solutions capable of significantly improving the predictiveaccuracy of rare cases in forecasting tasks using imbalanced timeseries data. We extend the application of resampling strategiesto the time series context and introduce the concept of temporaland relevance bias in the case selection process of such strategies,presenting new proposals. We evaluate the results of standardregression tools and the use of resampling strategies, with andwithout bias over 24 time series data sets from 6 different sources.Results show a significant increase in predictive accuracy of rarecases associated with the use of resampling strategies, and theuse of biased strategies further increases accuracy over the non-biased strategies.