Time-Series Classification for Dynamic Strategies in Multi-Step Forecasting

TL;DR

The paper tackles the lack of a universally optimal forecasting strategy for multi-step time-series forecasting by introducing DyStrat, a framework that treats dynamic strategy selection as a time-series classification problem. By constructing a pool of candidate MSF strategies and learning a policy to predict the locally optimal strategy for each instance, DyStrat achieves substantial improvements over fixed strategies across diverse datasets and horizons, with strong generalization to unseen data. The approach yields notable reductions in MSE and higher top-1 accuracy in strategy selection, suggesting a practical path to more accurate long-horizon forecasts in real-world domains. This dynamic, instance-level strategy selection has broad implications for MSF tasks, offering robustness across data scales, model complexities, and horizon lengths.

Abstract

Multi-step forecasting (MSF) in time-series, the ability to make predictions multiple time steps into the future, is fundamental to almost all temporal domains. To make such forecasts, one must assume the recursive complexity of the temporal dynamics. Such assumptions are referred to as the forecasting strategy used to train a predictive model. Previous work shows that it is not clear which forecasting strategy is optimal a priori to evaluating on unseen data. Furthermore, current approaches to MSF use a single (fixed) forecasting strategy. In this paper, we characterise the instance-level variance of optimal forecasting strategies and propose Dynamic Strategies (DyStrat) for MSF. We experiment using 10 datasets from different scales, domains, and lengths of multi-step horizons. When using a random-forest-based classifier, DyStrat outperforms the best fixed strategy, which is not knowable a priori, 94% of the time, with an average reduction in mean-squared error of 11%. Our approach typically triples the top-1 accuracy compared to current approaches. Notably, we show DyStrat generalises well for any MSF task.

Figures (7)

• Figure 1: Current approaches to multi-step forecasting do not consider the temporal dynamics of locally optimal strategies. Therefore, they make predictions using a single forecasting strategy, such as direct, dirrec, and rectify. An example time-series (Mackey-Glass) is shown with high variance in locally optimal strategies (a), colour of dots show which example strategy performs optimally. We compare our method, DyStrat, to these current approaches using the example strategies and achieve two-fold better local optimality (b). The error, using 1000 test points and 10 repeats, of each approach shows consistent significant error reduction using DyStrat (c).
• Figure 2: The top-1 accuracy (the proportion of within-task instances where a strategy is optimal) aggregated over all datasets and task settings. DIRMO and RECMO include all $\sigma$ parameters.
• Figure 3: Raw MSE from \ref{['alloverL80']} is mean-averaged and ranked per-task for each dataset (a). Raw MSE from \ref{['alloverL80']} is ranked per-instance and mean-averaged for each dataset (b). DS$C$ is DyStrat using $C$ as the classifier.
• Figure 4: Raw MSE from \ref{['alloverH160']} is mean-averaged and ranked per-task for each dataset (a). Raw MSE from \ref{['alloverH160']} is ranked per-instance and mean-averaged for each dataset (b). DS$C$ is DyStrat using $C$ as the classifier.
• Figure 5: DyStrat is given all possible candidate strategies (brute force in blue) and a restricted set that excludes recursive strategies (green). We sample strategies uniformly and plot their relative error to the optimal dynamic strategy. Lines show medians and shaded regions show interquartile range across 30 runs.