Exploring the association between time series features and forecasting by temporal aggregation using machine learning

When a forecast of the total value over several time periods ahead is required, forecasters are presented with two temporal aggregation (TA) approaches to produce required forecasts: i) aggregated (AF) or ii) aggregate data using non-overlapping (AD). Often, recommendation frequency relevant decision eventual will support and then forecast. However, this might not be always best choice we argue that both AF AD may outperform each other in different situations. Moreover, there lack evidence on what indicators determine superiority approach. We design execute an empirical experiment framework first explore performance these monthly series M4 competition dataset. further turn problem into classification supervised learning by constructing database consisting features as predictor model class labelled AF/AD response/outcome. build machine algorithms investigate association between AD. Our findings suggest consistently generate accurate results for every individual series. shown significantly better than series, especially longer horizons. set extracted input predict accurately whether should used. find out Random Forest (RF) most approach correctly classifying outcome assessed statistical measures such misclassification error, F-statistics, area under curve, utility measure. The RF reveals curvature, nonlinearity, seas_pacf, unitroot_pp, mean, ARCHM.LM, Coefficient Variation, stability, linearity, max_level_shif among important driving predictions model. indicate strength trend, ARCH.LM, hurst, autocorrelation lag 1, seas_pacf favour approach, while lumpiness, entropy, seasonality increase chance performing better. conclude study summarising present agenda research.