Towards Econometric World Models: Structural Multimodality and the TERRASTAT Dataset

An econometric world model should synthesize the millions of time series recording economic activity into a tool that forecasts, simulates scenarios, and reports its outputs in interpretable quantities. First, we introduce **Terrastat**, a corpus of one billion economic time series from FRED, Eurostat, and the OECD. The corpus is multimodal: each series carries text such as its title, units, and methodological notes. **Terrastat** is temporally aligned and can serve as a massive multivariate panel, updated in real time at varying frequencies. Second, we argue for **structural multimodality**: multimodal information (numeric history and text) is used to set and control an explicit econometric state---level, trend, and seasonality---rather than an end-to-end black-box forecast. By construction, this is the structure that macroeconomic agencies and policymakers already reason over, audit, and manipulate in their reports and decisions. We propose this structure itself as a symbolic interface for interacting with the model, usable by human analysts and AI agents alike. We implement a univariate-panel version of structural multimodality in **NeuroSym-CBF**, trained on the monthly FRED subset, FRED-194k. In addition to a structural time series model head, **NeuroSym-CBF** routes numerical and text embeddings through a sparse Top-$K$ bottleneck whose codes act on the state's coefficients, increasing robustness against the causal-reliability gap identified in concept bottlenecks. **NeuroSym-CBF** attains lower mean scaled MAE than Chronos-2, Chronos-Bolt, TimesFM-200M, and AutoETS at all five context lengths tested, improves $6.7$--$12.6$% over an ablation that keeps the same backbone and structural head but removes the multimodal pathway, and produces smaller 90th-percentile errors than every baseline, indicating better performance on difficult forecasting cases. Intervening on a single bottleneck coordinate produces a measurable forecast change that can be traced to level, trend, and seasonal coefficients. These results provide an initial step towards econometric world models with explicit, inspectable, and intervenable states.