An e-commerce demand simulator that explicitly tracks merchant actions and external events posted the lowest reported error scores in the study’s Alibaba 1688 tests at the next-5 and next-10 settings, and stayed below two named baselines in longer rollouts. In a separate online deployment, the group using the model to plan budgets recorded a reported 13% increase in lifetime value (LTV) and a 15% average improvement in store-level return on investment compared with the control group. But the online section does not report treatment and control group sizes, the randomization procedure, uncertainty intervals or significance tests. The figures are therefore reported comparisons, not established causal effects.
A forecasting model built around decisions
Called CEDAR, the system is a two-stage framework. An Action-Interleaved Transformer models state transitions alongside merchant actions, while a Residual Correction Module uses representations of aligned external-event signals. The evaluation used over 32 million Alibaba 1688 product trajectories with paired state-action sequences and those event signals. The experiments focused on four questions: comparative simulation performance, the value of explicit action modeling, residual correction for outside impacts, and online budget-planning effects.
The offline data covered approximately 32 million product trajectories from 2024 and 2025. The researchers used overlapping 15-week windows, held out the final window of 2025 for testing, and used preceding windows for training. All comparison models used identical data splits, input horizons and prediction windows. The reported scores were means and standard deviations from five independent runs with different random seeds.
Lower errors, even as the horizon stretches
On Alibaba 1688, CEDAR had the lowest reported MSE, MAE and NMSE—the study’s three error measures—in both the next-10 and next-5 settings. In the next-5 setting, MSE was 0.182 ± 0.006, MAE was 0.0603 ± 0.001 and NMSE was 0.0830 ± 0.001. Because the figures are summarized over five runs, the ± values describe run-to-run variation; the supplied analysis does not report formal significance tests for the offline comparisons.
CEDAR’s MSE rose as the forecast reached further ahead, but it remained below PatchTST and PETFormer at next-15, next-20 and next-25. With 10 weeks of historical observations, its MSE increased from 0.182 at next-5 to 1.612 at next-25. At that longest horizon, PatchTST scored 2.847 and PETFormer 2.561, compared with CEDAR’s 1.612. That pattern shows the model’s performance within this evaluation; it does not establish superiority across all datasets, product categories or marketplaces.
Actions and event corrections mattered
The action-modality analysis offered a qualitative look at how forecasts responded to interventions. In a representative product example, CEDAR was described as more responsive to changing advertising plans and as capturing a traffic decline after advertising expenditure was reduced. Covariate-based baselines showed limited sensitivity to those intervention signals. The analysis is partly qualitative, so this evidence illustrates trajectory response rather than supplying a general effect estimate.
An ablation comparison pointed in the same direction. The full model’s next-10 MSE was 0.414 and MAE was 0.132. With one-stage training, the scores were 0.471 and 0.158; an Action-Interleaved Transformer-only variant recorded 0.489 and 0.177. The pattern is consistent with the authors’ interpretation that the staged design and residual correction both mattered, although no inferential uncertainty was reported for these comparisons.
A second dataset broadens the test, but not the claim
CEDAR also performed better on the public Kaggle Store Sales dataset under next-5 testing, with lower MSE, MAE and NMSE than PatchTST and PETFormer. Its MSE was 0.5819, versus 0.6814 for PatchTST and 0.6321 for PETFormer. MAE was 0.3680 versus 0.4112 and 0.4623, while NMSE was 0.3778 versus 0.4425 and 0.4104.
That result broadens the forecasting comparison, but not the counterfactual budget-planning claim. The Kaggle dataset lacks explicit merchant actions, so it does not reproduce the intervention-simulation task tested with the Alibaba data.
The online result is the most consequential—and least documented
The reported Alibaba 1688 deployment ran from Jan. 1 through Jan. 30, 2026. It initially involved 239 cooperative merchants and 245 prediction-service orders; the report also lists 8.77 million RMB in transaction value and an initial repurchase rate of 61%. The CEDAR treatment group had a reported 13% increase in LTV over control, while store-level ROI improved 15% on average.
Those figures are useful as an outcome signal, but they cannot settle whether the model caused the gains. Overall deployment counts are given, but the report does not provide treatment-control sample sizes, allocation ratio, randomization details, confidence intervals or significance tests. Without those details, the online results should be read as a reported comparison between groups rather than a statistically established causal effect.
What can be checked next
The paper is identified in the supplied record as arXiv:2608.25871v1, dated 26 August 2026, and is a preprint. The authors say they are actively working toward releasing a partially anonymized E-Comm 15-Week dataset, but no current access details are reported. The reported training cost was 154 minutes for Stage I plus 80 minutes for Stage II, about 234 minutes total; generating event embeddings took about four hours as a one-time offline cost.
For transparency, the acknowledgments name the National Natural Science Foundation of China, the Natural Science Foundation of Anhui Province, the New Generation Artificial Intelligence-National Science and Technology Major Project, the CCF-1688 Yuanbao Cooperation Fund, the National Key R&D Program of China, the Guangdong Basic and Applied Basic Research Foundation and the Key-Area Special Project of Guangdong Provincial Ordinary Universities as funders.
Paper data and sources
Original title: CEDAR: Controlled and Event-Driven Demand Forecasting via Residual Decomposition
Authors: Junjie Meng, Ranxu Zhang, Zi-an Zhang et al.
Journal/Repository: Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2 (KDD '26), August 09-13, 2026, Jeju Island, Republic of Korea. ACM, 2026, 12 pages
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: 10.1145/3770855.3818338
Original paper · Full text