STARRY-NET: A Neuro-Symbolic Model for Time Series Forecasting

Forecasting time series in domains such as macroeconomics, tourism, public health, and retail demand requires reasoning over rich structural dynamics with stochastic uncertainty, including evolving growth rates, seasonal cycles, and regime shifts. However, such dynamics are often lacking in black-box foundational models. While classical state-space models provide the rigorous interpretability required for high-stakes decision-making, they face three fundamental challenges: (1) isolated per-series estimation cannot leverage shared statistical regularities across extensive yet diverse panels; and (2) rigid analytical equations are vulnerable to data-level shocks and misspecifications; and (3) inference time parameter optimizations (e.g., MLE or MCMC) are slow and expensive. In this paper, we introduce Starry-Net, a neuro-symbolic framework for auditable forecasting. Given any structural time series model, Starry-Net uses a neural backbone to estimate its parameters globally across diverse corpora, while an end-to-end orthogonal neural remainder absorbs and predicts complex residual dynamics without compromising structural identifiability and interpretability. We instantiate this framework with a structural family based on generalized exponential smoothing (ETS/TBATS), a gold standard in industrial forecasting refined across decades of practice. Across real-world benchmarks, Starry-Net substantially outperforms classical and TSFM counterparts in accuracy and computational efficiency---all while providing transparent, policy-intervenable forecasts. By offloading dominant structural patterns to symbolic decoders, the architecture achieves high adaptivity in small panel fine tuning, which minimizes unwanted transfer bias and data leakage from off-the-shelf black-box models.