Organizations: School of Artificial Intelligence and Data Science, University of Science and Technology of China, Hefei, China · Alibaba Group, Hangzhou, China · Thrust of Artificial Intelligence, The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, China · Department of Computer Science and Engineering, The Hong Kong University of Science and Technology, Hong Kong SAR, China
Forecasting in large-scale e-commerce marketplaces is increasingly required to support planning: merchants need to evaluate sales outcomes under future action sequences such as budget schedules, rather than passively predicting what happens next. However, most existing time series forecasting (TSF) approaches remain inherently passive. Even when incorporating operational decisions as auxiliary covariates, they typically optimize for correlation-based extrapolation under historical policies. This design suffers from autoregressive inertia and conflates endogenous market evolution with decision-induced transitions, leading to policy-insensitive rollouts and unreliable counterfactual analysis. To bridge this gap, we propose CEDAR (Controlled and Event-Driven Demand forecasting via Action-aware Residual decomposition), a two-stage framework for robust decision-conditioned simulation. In Stage I, an Action-Interleaved Transformer learns controllable action-conditioned state transitions for rollout under planned interventions. In Stage II, a Residual Correction Module leverages external event signals and LLM-assisted text representations to align noisy event descriptions with product context and correct event-driven deviations. Our study is enabled by a large-scale real-world dataset from Alibaba 1688, comprising approximately 32 million product trajectories with paired state-action sequences and aligned event signals. Extensive offline experiments and online controlled experiments in production demonstrate that CEDAR consistently improves simulation accuracy over strong TSF baselines and delivers practical gains for real-world budget planning.
Figures & tables
Figure 1. Performance comparison of four TSF baselines. All models are trained on the E-Com 15-Week dataset for 10 steps prediction.
Figure 2. Framework of CEDAR. The AIT module is trained in the first stage, and frozen in the second stage.
Metric
Setting
CEDAR
Informer
TFT
PatchTST
PETFormer
Timer-XL
MSE ↓
next 10
0.414 ± 0.015
32.2 ± 21
0.722 ± 0.024
0.849 ± 0.026
0.743 ± 0.021
2.98 ± 0.14
next 5
0.182 ± 0.006
3.44 ± 1.8
0.572 ± 0.015
0.424 ± 0.011
0.434 ± 0.013
1.34 ± 0.05
MAE ↓
next 10
0.132 ± 0.002
3.23 ± 2.1
0.194 ± 0.007
0.201 ± 0.006
0.192 ± 0.006
0.627 ± 0.035
next 5
0.0603 ± 0.001
0.730 ± 0.35
0.175 ± 0.005
0.129 ± 0.004
0.139 ± 0.004
0.472 ± 0.018
NMSE ↓
next 10
0.189 ± 0.004
14.7 ± 9.2
0.337 ± 0.012
0.387 ± 0.013
0.339 ± 0.011
1.36 ± 0.08
next 5
0.0830 ± 0.001
1.57 ± 0.80
0.267 ± 0.009
0.191 ± 0.006
0.198 ± 0.007
0.612 ± 0.022
Table 1. Performance comparison of CEDAR and other baselines. The next 5 setting predicts the next 5 weeks given the previous 10 weeks, while the next 10 setting predicts the next 10 weeks given the previous 5 weeks. All offline results are reported as mean ± standard deviation over five independent runs.
Ablation Variants
MSE ↓
MAE ↓
next 10
next 5
next 10
next 5
Full Model
0.414
0.182
0.132
0.0603
w/o Recent states
0.443
0.209
0.137
0.0634
w/o Product metadata
0.466
0.227
0.143
0.0654
w/o AIT Prediction
0.499
0.264
0.181
0.0697
w/o Holiday keywords
0.451
0.213
0.169
0.0651
Table 2. Ablation study results using MSE and MAE metrics. One-stage training refers to jointly training the Action-Interleaved Transformer and the Residual Correction Module.
Figure 3. Visualization of predicted views trajectories of different baselines.
Figure 4. Visualization of predicted views trajectories with next-1 setting of different baselines.
Metric
CEDAR
PatchTST
PETFormer
MSE ↓
0.5819
0.6814
0.6321
MAE ↓
0.3680
0.4112
0.4623
NMSE ↓
0.3778
0.4425
0.4104
Table 3. Performance comparison on the public Kaggle Store Sales dataset under the next-5 setting.
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
Component
Time / Epoch
Epochs
Total Time
Notes
CEDAR Stage I
≈ 11 min
14
154 min (2h 34m)
Base-stage training
CEDAR Stage II
≈ 16 min
5
80 min (1h 20m)
Residual correction training
Event embedding generation
–
–
≈ 4h
One-time offline cost
PatchTST
≈ 7 min
17
119 min (1h 59m)
Baseline reference
PETFormer
≈ 21 min
31
651 min (10h 51m)
Baseline reference
Appendix
Table 4. Wall-clock cost analysis of CEDAR and representative baselines. The LLM-based event embedding generation is a one-time offline preprocessing cost.
Horizon
CEDAR
PatchTST
PETFormer
next 5
0.182
0.424
0.434
next 10
0.414
0.849
0.743
next 15
0.734
1.526
1.382
next 20
1.103
2.173
1.945
next 25
1.612
2.847
2.561
Appendix
Table 5. Long-horizon rollout performance on Alibaba 1688 measured by MSE. All settings use 10 weeks of historical observations as input.
Demand forecasting increasingly requires combining two complementary sources of information: historical sales reveal recurring numerical dynamics, while future promotions, holidays, price changes, and platform interventions provide forward-looking knowledge. Existing text-enhanced forecasting methods often encode such context into generic representations and fuse it uniformly with time-series features, without explicitly distinguishing which semantic effects are forecast-relevant or how they should modify future dynamics. We introduce ReasonCast, a structured semantic intervention framework that translates event knowledge into forecast-specific operations. An agent examines the event context, the no-text forecast, and its uncertainty to determine whether textual reasoning is needed. Rather than injecting free-form text, ReasonCast represents event knowledge through structured fields describing event relevance, demand direction, temporal shape, amplitude, and peak intensity. These fields interact selectively with temporal components of a time-series foundation model. An additive path corrects local trends and temporal shapes, while a multiplicative path captures event-driven level shifts. ReasonCast introduces a forecast-grounded post-training curriculum. Schema SFT establishes semantic fields; semantic-field RL calibrates direction, shape, amplitude, and peak judgments; and forecast-utility RL evaluates semantic interventions through a frozen forecaster, aligning reasoning outputs with marginal forecast improvement. ReasonCast lowers WMAPE by 3.29, 1.25, and 0.47 percentage points on holiday-sensitive categories, mega-sale-sensitive categories, and M5 event windows, respectively. On stable-sales periods, indiscriminate semantic intervention increases WMAPE by 1.68 percentage points, whereas suppressing unnecessary intervention preserves the numerical backbone.
Deploying a new decision policy creates a cold-start problem for prediction models whose targets depend on the policy's actions: historical observations reflect earlier policies, while real observations under the new policy are not yet available. Simulation offers a way to address this gap by rolling out the target policy across counterfactual scenarios and using the resulting trajectories to learn how the system responds to those controls. The simulation-to-reality (Sim2Real) transfer of this simulator-trained model can then be backtested by evaluating it against real observations from past deployments. Using two real-world inventory-control deployments, we evaluate this process from three angles: simulator fidelity, zero-shot transfer to real behavior, and adaptation as real target-policy observations accumulate. The simulator-trained forecaster achieves lower point-estimate mean absolute percentage error (MAPE) than the same architecture trained on historical real data, reducing MAPE by 1.2-3.1 percentage points in Study 1 and 12.5-18.7 points in Study 2. After deployment, lightweight calibration using early real observations further reduces error by up to 2.5 percentage points. These results provide empirical evidence that simulator-generated counterfactual data can support cold-start forecasting under a new policy, and the resulting model can be further refined as real deployment data become available.
Time series foundation models (TSFMs) are pretrained on series from diverse domains, where demand series make up only a small fraction. Demand data has properties that such corpora rarely contain: Short histories, frequent zeros, censoring by stock-outs, and exogenous events that the series does not record. To this end, we propose EXAONE Demand, built on 1) a demand-specific corpus and 2) a demand-aware adapter. For the corpus, we assemble 11.3M series and 48.4B observations from 73 sources, and a synthetic generator supplies the behaviour that open demand data under-represents. For the adapter, we attach low-rank branches to a frozen general-domain backbone, one for each of the four demand classes (smooth, intermittent, erratic, and lumpy), and a router that reads eight scale-free statistics of the input series decides how much each branch contributes. We build EXAONE Demand in two versions, one trained on real-world and synthetic demand together and one trained on the synthetic corpus alone. On 22 held-out datasets, both versions outperform 36 TSFMs, and real-world demand adds a gain over synthetic data alone.