JD's TimeHF 2.0 puts regional sales, structured conditions and business notes into one forecast, lifting national daily accuracy 18.25% over the production baseline
Synopsis
JD's supply-chain technology team released TimeHF 2.0, which conditions a single forecasting process on national and distribution-center sales histories, structured covariates and users' natural-language plans, and outputs both daily sales and cumulative quantities for procurement and replenishment windows. In an online parallel evaluation over 200,000 SKUs, national daily and planning-window accuracy improved by 18.25% and 17.94% over the original production deep learning baseline, with 19.44% and 18.99% at distribution-center level.
Interpretation
Retail supply-chain forecasting no longer extrapolates from a single item's history alone; national trajectories, distribution-center trajectories, available covariates and plans written by business staff jointly condition the prediction, which yields daily sales and planning-window totals together. The earlier TimeHF 1.0 supported item-level planning through temporal modeling and human-feedback-guided optimization, mapping historical sequences to future sales; coordinating procurement with regional inventory and replenishment required joint, context-aware forecasts. In JD's online parallel evaluation of 200,000 SKUs from January to September 2026, TimeHF 2.0 improved national daily and planning-window accuracy by 18.25% and 17.94% over the original production deep learning baseline, against 13.84% and 13.27% for TimeHF 1.0.
The model first encodes each sales target's history and covariates, then lets language-model representations of business descriptions query that target's numerical memory, so a campaign note lands on a specific product and date. Language-model representations are connected to numerical evidence through a semantic alignment module rather than concatenated as extra features. The encoder applies attention along the time axis and the variable axis separately; a regional branch reads national and peer distribution-center sales for the same product, a covariate branch reads campaign schedules and already-issued weather forecasts, and the decoder uses text-conditioned future queries with shared and routed experts to output a sales distribution over ordered bins.
Training constrains both the daily distribution and the planning-window total, so daily deviations do not quietly accumulate into wrong replenishment quantities. With continuous regression heads alone, national planning-window improvement was 15.02% for MSE, 15.86% for MAE and 16.34% for Huber, and 17.34%, 17.75% and 17.99% at distribution-center level. An ordered-bin head with cross-entropy and ranked probability score raised national and distribution-center results to 17.12% and 18.54%, and adding the regression penalty and cumulative supervision reached 17.94% and 18.99%.
Perspective
The evaluation is limited to online parallel forecasting for 200,000 SKUs from JD's retail catalogue during January–September 2026, with training on 912 billion data points sampled from retail records spanning 2023–2025, compared against the original production deep learning baseline and TimeHF 1.0 on matched product–region–origin records.
Sparse histories, stock-constrained observations, ambiguous plans and abrupt local peaks still affect what can be learned; the report notes that a close accumulated quantity can coexist with substantial local errors. The native multimodal-to-multimodal explanations envisioned for TimeHF 3.0 remain a future direction and are not evaluated here.
