Structure at the Top, Noise at the Leaves: Hierarchy-Aware Residual Adaptation of Frozen Time-Series Foundation Models

Time-series foundation models (TSFMs) forecast each series on its own and ignore the aggregation structure that ties many real systems together. We ask whether an explicit hierarchy can improve a frozen TSFM without touching its weights. This work continues the hierarchical residential load forecasting study of Sucaldito & Ostia (2025), which found that per-series AutoARIMA and LightGBM models forecast aggregates well but degrade sharply on zero-inflated appliance series, and which called for hybrid, hierarchy-aware models. We introduce such a hybrid, the Hierarchy-Aware Residual Adapter (HAA), a 19k-parameter MLP that sits on top of frozen TimesFM 2.5 forecasts. It predicts leaf-level residuals from structural context: bottom-up category, household and system aggregates, the leaf's share of each, recent temporal statistics, and category identity. Bottom-up aggregation of the corrected leaves keeps every level exactly coherent. On a four-level residential energy hierarchy (Total → 25 households → 105 categories → 218 sub-metered appliances, Pecan Street New York, 15-minute resolution), we evaluate a 24-hour forecast against naïve, seasonal-naïve, Chronos and TimesFM baselines. HAA lowers pooled RMSE relative to frozen TimesFM at every level: by 2.0% (appliance), 1.1% (category), 5.1% (household) and 20.2% (system total; Diebold–Mariano $p<10^{-6}$). A closer analysis gives a more cautious picture. HAA beats TimesFM on only 24%, 30% and 36% of individual appliance, category and household series. Series-bootstrap confidence intervals for the lower-level gains include zero. Almost all of the gain at the total level comes from removing a systematic under-forecast (the bias share of TimesFM's total-level MSE falls from 73% to 47%), while the error variance actually grows. The adapter also produces negative loads in 31% of leaf predictions and worsens MAE and MAPE at the leaves. Projecting leaf predictions onto the non-negative orthant before aggregation raises the total-level RMSE gain to 29.4%. We argue that structural context is most useful as an aggregation-amplified bias corrector. We also release a protocol (rolling origins, household hold-out, hierarchy permutation, cross-city transfer) for testing whether such adapters show real structural generalization.