Organizations: Supply Chain Tech Team Y, JD.com, China · Tsinghua University, China · Beijing Jiaotong University, China · The University of Hong Kong, China
Large-scale retail supply chains need sales forecasts for specified products, future dates or planning windows, and regions under the conditions available at the forecast origin. Regional sparsity, changing environments, volatile sales, and unfamiliar business situations make these dimensions difficult to model together. We introduce TimeHF 2.0, a multimodal, multivariate, and multi-region collaborative forecasting framework that makes flexibly expressed business knowledge usable in numerical prediction. A coverage-aware directed encoder integrates regional sales and applicable covariates. A semantic alignment module transfers language-model hidden representations of natural-language plans and domain experience into target-specific forecasting context. Learning against sales observations connects these meanings to numerical evidence, allowing business users to supply experience, future plans, and changing circumstances without defining a new numerical feature for each description. Text-conditioned future queries and shared/routed experts adapt prediction to the supplied context. An ordered-distribution head allocates fine resolution to routine fluctuations and coarse resolution to extreme sales movements, while daily/window learning connects forecasts to cumulative planning quantities. Online parallel forecasting at JD yields relative accuracy improvements of 18.25% and 17.94% for national daily and planning-window sales, and 19.44% and 18.99% for DC daily and planning-window sales, over the original production deep learning baseline. Component comparisons, loss-design results, and product-type analyses examine the roles of regional information, covariates, business knowledge, context-dependent processing, and robust planning-oriented learning.
Figures & tables
Figure 1: Supply-chain forecasting context and the TimeHF roadmap. (a) Solid histories and dashed future trajectories illustrate national/DC forecasts; daily paths and window totals support procurement, inventory placement, replenishment, and fulfillment. The vertical line marks the forecast origin. (b) Each generation connects a business need to required capabilities and a technical route. TimeHF 2.0 is the present framework, connecting regional numerical evidence with language-conditioned forecasting. TimeHF 3.0 envisions native multimodal inputs and outputs, combining forecast accuracy with native explanations. The TimeHF 2.0 route depicts multiple numerical inputs and a single target sales trajectory; the model is applied to national and DC targets. Routes and trajectories are schematic; the first generation’s sequence-to-sequence label describes its input/output mapping.
Figure 2: Business situations and forecasting challenges in TimeHF 2.0. The top strip states the shared product–time–region forecasting task under available covariate and business context. C1–C4 illustrate regional histories, condition scope and availability, product types and seasonal/activity regimes, and daily/window quantities, including fine routine and coarse extreme variation. C5 connects user-expressed plans and domain experience with unfamiliar situations and scarce comparable evidence. Response labels summarize the corresponding components in Section 3 . Temp denotes an origin-visible temperature forecast. PV denotes observed page views. Conditions, histories, and response paths are schematic examples, not measured forecasts.
Figure 3: TimeHF 2.0 framework and section correspondence. Section 3.2 prepares time-series inputs and coverage context; Section 3.3 applies temporal attention, directed variable attention, and FFN to form Hi and Cg . Section 3.4 transfers language-model hidden representations of users’ plans and business knowledge through semantic alignment with numerical memory into context ci,h . Section 3.5 applies that context in future queries, reads the original Hi , and routes the text-conditioned decoding state through shared/routed expert processing and a shared head for daily/window forecasts. The dashed supervision path corresponds to Section 3.6 ; r and Cg ports denote regional and covariate context. Token grids and expert/probability symbols are schematic.
Relative accuracy improvement (%)
Model
Daily
Planning window
Original DL baseline
—
—
TimeHF 1.0
13.84
13.27
TimeHF 2.0
18.25
17.94
Table 1: National online forecasting results. Entries are relative accuracy improvement (%).
Relative accuracy improvement (%)
Model / variant
Daily
Planning window
Original DL baseline
—
—
TimeHF 1.0
12.02
11.56
TimeHF 2.0
19.44
18.99
w/o covariate information
17.34
16.03
w/o cumulative supervision
17.71
16.50
Table 2: DC online results and component ablations. Entries are relative accuracy improvement (%).
Figure 4: Online retail sales fluctuations around a routine baseline of 100. Sales are indexed as 100×yt/b , where b is the routine baseline. A1 denotes a one-off large order from an individual retail customer; A2, occasional repeated bulk purchases; and A3, uncontrolled order concentration in private customer channels. These isolated peaks have no matching campaign, PV, or weather signal. In contrast, the shaded promotion has an available campaign signal and elevated PV. Learning should preserve this supported lift while limiting the influence of isolated extreme amplitudes.
Window relative accuracy improvement (%)
Output / supervision
National
DC
Continuous regression: MSE
15.02
17.34
Continuous regression: MAE
15.86
17.75
Continuous regression: Huber
16.34
17.99
Ordered-bin head: CE + RPS
17.12
18.54
Ordered-bin head: full objective
17.94
18.99
Table 3: Output-head and training-objective comparison. Entries are planning-window relative accuracy improvement (%).
National
DC
Product type
Baseline
TimeHF 1.0
TimeHF 2.0
Baseline
TimeHF 1.0
TimeHF 2.0
Bestsellers
—
16.23
19.52
—
13.44
20.11
Seasonal
—
12.16
17.64
—
9.45
24.52
Long-tail
—
15.77
20.15
—
12.64
28.66
New products
—
3.01
9.28
—
2.19
11.25
Table 4: Planning-window relative accuracy improvement (%) across product types. Baseline columns refer to the original production deep learning model.
Figure 5: Online forecasting case with an added campaign plan. Shading marks the main activity; forecast revisions affect the sustained level without reproducing daily peaks.
Figure 6: Online forecasting case with a revised campaign schedule. Gray and orange shading mark the earlier and revised periods; level and timing differences remain visible.
Figure 7: Online forecasting case with occasion knowledge. The main peaks align, with different amplitudes and an earlier forecast shoulder.
Figure 8: Online local-quantity comparison; the late rebound remains underestimated.
Figure 9: Online forecasting case with a locally anchored forecast and a restrained regional lift.
Figure 10: Selective regional collaboration. The forecast retains the local sales level while avoiding the growth pattern of unsuitable national and northern references; South C supplies the applicable regional context.
Time series forecasting predicts future values from ordered observations using statistical, neural, and probabilistic models to support decisions under uncertainty.
Time-Series Foundation Models (TSFMs) excel at zero-shot unimodal forecasting using numerical data, but unlike LLMs they cannot consume multimodal, non-numerical context that often shape real-world trajectories. In this work, we bridge this gap and argue for a multimodal time-series forecasting approach that post-trains LLMs to act as context-guided revisors over strong numerical TSFM priors. We introduce PostTime, a post-training recipe combining Supervised Fine-Tuning (SFT) and Reinforcement Learning with Verifiable Rewards (RLVR), along with a methodology to generate automated reasoning traces for forecast revisions. PostTime teaches an LLM to generate context-conditioned forecast interventions -- decisions to revise, preserve, or ignore the TSFM prior based on the multimodal context. We evaluate this approach on the TimesX multimodal forecasting benchmark using a Gemma-3-4B LLM and TimesFM-2.5 TSFM, and show that it significantly outperforms standalone TSFMs, LLM-only baselines, and existing multimodal forecasting approaches.
Time series forecasting models often benefit from historical patterns. Inspired by Retrieval-Augmented Generation (RAG), recent research explored retrieving relevant historical time series segments to enhance forecasting. However, relying solely on time series similarity is often insufficient for retrieval under non-stationarity. To address this, we propose a multimodal approach: a Semantics-Enhanced Retrieval-Augmented Time Series Forecasting framework, SERAF. Unlike mainstream approaches that depend only on time series similarity, SERAF conducts dual retrieval over the time series and their self-generated textual descriptions. It retrieves two complementary sets of historical patterns and corresponding futures, which are selectively and jointly used to guide future predictions. Experiments across seven real-world datasets demonstrate the effectiveness of SERAF in bridging numerical and semantic views of time series compared with state-of-the-art baselines.
Shiqiao Zhou, Zipeng Wu, Holger Schöner +3
School of Computer Science, University of Birmingham, Birmingham, United Kingdom · School of Mathematics, University of Birmingham, Birmingham, United Kingdom · Siemens AG, Munich, Germany +1
Cross-domain multimodal time series forecasting is a challenging task, requiring models to integrate precise numerical comprehension, cross-domain semantic understanding, and effective multimodal fusion. Existing approaches either build Time Series Foundation Models (TSFMs) from scratch or leverage pretrained Large Language Models (LLMs). However, TSFMs often overlook semantic understanding and lack the ability to perform future-oriented semantic reasoning, and LLMs struggle with numerical comprehension and accurate quantitative forecasting. To overcome these limitations, we propose KairosAgent, a novel agentic framework for multimodal time series forecasting, including an LLM-based reasoner and a TSFM-based forecaster. KairosAgent unifies textual reasoning and numerical forecasting by dynamically invoking analytical tools to enhance the numerical understanding and semantic reasoning capabilities of LLMs. The reasoning results are subsequently fused into the TSFM pipeline, enabling more accurate and reliable future predictions. To further improve the reasoning, we curate a large-scale corpus of high-quality trajectories, alongside a reinforcement learning from forecasting paradigm with multi-turn refinement and turn-level credit assignment. Experiments demonstrate that KairosAgent achieves superior zero-shot forecasting performance while maximizing the utility of pretrained LLMs and TSFMs, presenting a promising direction for efficient and interpretable time series agents. The project page is at https://foundation-model-research.github.io/KairosAgent .