GLoR-forecast: state-contingent global-local text routing for financial and socio-economic time-series forecasting

Abstract

Financial and socio-economic forecasting often requires information beyond numerical histories, as policy statements, news reports, institutional releases, and event descriptions may signal structural regimes or short-term shocks. Recent multimodal time-series methods incorporate timestamp-aligned texts as auxiliary variables in numerical forecasting backbones, but typically treat all textual signals as homogeneous local covariates. This overlooks an important distinction between local event-level text and broader contextual narratives. This paper proposes GLoR-Forecast (Global-Local Text Routing for Forecasting), a state-contingent textual fusion framework for multimodal time-series forecasting. The model encodes timestamp-aligned texts with a frozen language model, projects them into compact local textual variables, and attentively aggregates them into a global contextual representation. A text router then uses the current numerical time-series state as a query to dynamically weight local and global textual features. The fused representation augments the numerical sequence and can be used with standard forecasting backbones; we instantiate it with iTransformer. Experiments on nine Time-MMD domains show consistent gains over TaTS, reducing average MSE by 5.8% and MAE by 4.7%. The strongest improvements appear in Economy, Social Good, Traffic, and Agriculture, suggesting that state-dependent textual routing helps exploit both slow-moving structural context and short-term event shocks.