AI-Enhanced Grouped Time-Series Forecasting with Weather and Holiday Effects

Abstract

Operational wholesale forecasting depends on two levers: external drivers such as weather and holidays, and hierarchical coherence across customer groups. Their interaction with modern deep sequence models on multi-year business-to-business data is rarely measured with formal uncertainty quantification. We benchmark four model families (error-trend-seasonal exponential smoothing (ETS), SARIMA, the Temporal Fusion Transformer (TFT), and the Time-series Dense Encoder (TiDE)) on 179 weeks from a Hungarian wholesale distributor (~5.1 M records, 93,287 modelled customer-item series), under a matched rolling-origin protocol with eight non-overlapping 12-week windows, and compare Bottom-Up against Top-Down reconciliation. Pooled across the windows, TFT significantly outperforms ETS, SARIMA, and TiDE at every horizon (group-clustered paired tests, $p \leq 0.007$) and attains the best macro-mean MAE and WAPE, though individual-window means can favour the statistical baselines. TFT also beats a covariate-equipped LightGBM reference $(p \leq 0.003)$, while the same covariates significantly worsen a per-series SARIMAX, and all families clear a seasonal-naive floor by a wide margin. Covariate effects are marginal on average (at most $\approx \mathbf{0. 6 \%}$ of MAE, even under perfect prognosis), yet regime-conditional: corrected holiday features cut group-total error by 5% on major-holiday peak weeks. Top-Down reconciliation degrades series MAE by 35-60%, so Bottom-Up is the default, under which the statistical baselines give the most accurate group totals in most windows. Per-group difficulty spans an order of magnitude, so TFT is the strongest global default but monitoring stays segment-aware.