Context Routing Benefits in Time Series Forecasting Identified

Researchers conducted a systematic study to pinpoint the specific conditions under which auxiliary context truly enhances multi-modal time series forecasting, rather than merely reflecting architectural side effects. The study identified two essential dataset-level prerequisites for context to be beneficial: first, the target variable must not be overwhelmingly predictable by its last value (i.e., exhibit low autocorrelation); and second, the auxiliary context must carry information about the target that is not already present in its historical data (i.e., possess non-zero conditional mutual information). Through extensive experiments using a 14.3B-parameter mixture-of-experts model (MoME) and other fusion mechanisms, the study found that text-conditioned expert modulation significantly reduces Mean Squared Error (MSE) only when both conditions are met. If either condition fails, the contribution of context routing diminishes to a baseline capacity, indicating no genuine signal from the context. Causal interventions, such as adding shortcuts or corrupting context quality, further validated these findings, demonstrating a direct link between these conditions and the effectiveness of context. The research provides a calibrated pre-training diagnostic to predict when context routing will be effective.