A Step in the Right Direction: Bridging Forecasting and Classification for Temporal Foundation Models

As temporal foundation models are developed to support a broader range of downstream tasks, a central question is whether forecasting performance translates into task-specific predictive utility. Directional prediction—whether a future observation increases relative to the latest observed value—is relevant to decision-making in many financial and operational settings. We study it as a shared task connecting time series forecasting and classification. We introduce a framework that reformulates forecasting datasets into horizon-specific classification tasks, enabling direct comparison between directions inferred from forecasting foundation models and those predicted by classification foundation models. Experiments across the six M4 Competition frequencies reveal contrasting horizon-dependent behaviour. On the Daily subset, directional accuracy derived from the forecasting foundation model Chronos-2 generally decreases with horizon, while that of the classification foundation model Mantis generally increases. A post hoc adjustment analysis further explores whether classifier-derived directional signals can complement point forecasts. Together, these findings establish directional prediction as a common ground for evaluating forecasting and classification approaches, motivating broader assessment of temporal foundation models beyond conventional forecast scores.