TL;DR
TimeHF 2.0 is a next-generation forecasting framework that integrates numerical time-series data with natural-language business insights (like promotional plans or domain expertise). By combining a coverage-aware regional encoder with LLM-based semantic alignment, it achieves an ~18-19% accuracy improvement over production baselines for JD.com’s massive retail supply chain.
What Problem Does This Paper Address?
Large-scale retail forecasting is notoriously difficult due to "cold starts" for new products, regional data sparsity, and extreme sales volatility. Traditional models rely heavily on historical numerical features, but they often miss the "why" behind the data. For instance, if a manager knows a specific "back-to-school" campaign is starting, that knowledge is often trapped in a text document or a human's head rather than a structured database.
The authors identify five core challenges (C1–C5) in this domain:
- Regional Heterogeneity: Local distribution centers (DCs) have different patterns than national averages.
- Variable Availability: Some data (like weather) is only known at the forecast origin, while other data (like page views) is historical.
- Demand Regimes: Products shift between stable, seasonal, and promotion-driven states.
- Planning vs. Precision: Supply chains need both daily accuracy and cumulative totals for procurement windows.
- Unfamiliar Situations: Models struggle when faced with new events that have no historical numerical precedent.
Methods
TimeHF 2.0 moves beyond simple sequence-to-sequence models by introducing a multimodal, multi-series architecture.
- Directed Encoder: Instead of a standard Transformer encoder, it uses "directed variable attention." This allows the model to selectively transfer patterns from national or peer-DC histories to a specific local target based on "coverage context" (e.g., service area size and city relevance).
- Semantic Alignment: This is the "secret sauce." The model uses a Large Language Model (LLM) to process natural-language descriptions of plans and experiences. A trainable alignment module then maps these linguistic representations into the same vector space as the numerical sales data.
- Text-Conditioned Decoder: The decoder uses a Mixture of Experts (MoE) approach. The "experts" (specialized sub-networks) are activated based on the context provided by the text. If the text describes a "promotion lead-in," the model routes the computation to experts trained for high-volatility regimes.
- Ordered-Distribution Head: Rather than predicting a single point estimate, the model predicts a probability distribution over "ordered bins." It uses fine-grained bins for routine fluctuations and coarse bins for extreme peaks, making the model more robust to outliers.
Datasets
The model was evaluated using JD.com’s industrial-scale retail data:
- Scale: 200,000 SKUs across hundreds of distribution centers.
- Training Corpus: 912 billion data points sampled from 2023–2025.
- Evaluation Period: January–September 2026 (online parallel forecasting).
- Inputs: Origin-visible sales histories, structured covariates (weather, page views), and natural-language business descriptions.
Key Insights and Ideas
- Language as a "Soft" Feature: The most striking idea is that text isn't just a label; it's a condition. By describing a campaign in prose, users can influence the forecast without needing to engineer a new numerical feature for every unique event.
- Robustness through Binning: By using non-uniform bins for the output, the model avoids being "pulled" too hard by unpredictable bulk orders (outliers) while still capturing the sustained "lift" from a planned promotion.
- Regional Collaboration: The model proves that "national" data is a powerful regularizer for "local" DC data, provided the model has a mechanism (the directed encoder) to ignore irrelevant regional signals.
Key Results
The framework was tested in a live "online parallel" setting at JD.com, comparing it against a deep learning production baseline:
- National Level: Relative accuracy improved by 18.25% for daily sales and 17.94% for planning windows.
- DC Level: Even higher gains, with 19.44% improvement for daily and 18.99% for windows.
- Product Types: The model was particularly effective for "Long-tail" and "Seasonal" products at the DC level, where data is usually sparsest.
- Ablation Studies: Removing the multimodal (text) component led to a measurable drop in performance, confirming that business knowledge provides a unique signal that numerical history cannot.
Limitations and Future Work
Despite the gains, the authors note that "abrupt local peaks" and "ambiguous plans" still pose challenges. For example, if a user provides a vague description of a plan, the model's response may be less accurate.
The roadmap for the future is TimeHF 3.0, which aims for "native multimodal-to-multimodal" modeling. The goal is to move beyond just taking text as input to providing native explanations—where the model doesn't just give a forecast but explains in natural language why it changed the prediction based on the provided context.
Summary updated