🤖 machine learning

CEDAR: Controlled and Event-Driven Demand Forecasting via Residual Decomposition

CEDAR is a two-stage framework that enhances demand forecasting for e-commerce planning by combining an Action-Interleaved Transformer for simulating policy-driven state transitions with a Residual Correction Module that leverages LLMs and event signals to address the limitations of passive time series models, thereby enabling robust counterfactual analysis and improved budget planning.

Original authors: Junjie Meng, Ranxu Zhang, Zi-an Zhang, Shujun Liu, Xiaoning Qi, Xiaozhou Xu, Yanyong Zhang, Hui Xiong, Chao Wang

Published 2026-08-27

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Junjie Meng, Ranxu Zhang, Zi-an Zhang, Shujun Liu, Xiaoning Qi, Xiaozhou Xu, Yanyong Zhang, Hui Xiong, Chao Wang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the vast, humming digital marketplace of modern e-commerce, a fundamental question has long puzzled planners: how do we know what will happen next? For decades, the standard answer relied on a passive form of prediction. Computers would study the past, looking at how sales moved in response to prices, holidays, and advertising, and then simply extend those lines forward. It was a method of extrapolation, assuming that the future would look much like the past, just slightly shifted. This approach worked well enough when the world was stable and merchants merely reacted to trends. But in the dynamic, high-stakes environment of global trade, merchants are not passive observers; they are active drivers. They constantly change their strategies, adjusting budgets, launching promotions, and altering prices to reshape demand before it even happens. The old way of forecasting, which treated these human decisions as mere background noise, began to fail. It could not answer the most critical question a business leader needs to ask: "If I change my plan today, what will the market look like tomorrow?"

This is the challenge that a team of researchers from the University of Science and Technology of China and Alibaba Group set out to solve. They recognized that to predict the future of a marketplace, one must stop treating the future as a fixed destination and start treating it as a path that can be steered. Their work, presented at a major data science conference, introduces a new system called CEDAR. Unlike traditional models that simply watch history unfold, CEDAR is designed to simulate what would happen if a merchant took specific actions. It does not just guess the next number; it calculates the outcome of a choice. By separating the natural flow of the market from the specific effects of human intervention, and by accounting for sudden, unpredictable external events, the researchers created a tool that can reliably test different budget strategies before a single dollar is spent.

The core of this new approach lies in a two-part architecture that mimics the way a merchant actually thinks and acts. The first part is a specialized engine that learns the direct link between a decision and its immediate result. In the past, models would mix a merchant's actions, like a discount or an ad spend, with the sales data, treating them as a single, jumbled stream of information. This caused the model to get confused, often assuming that a sales spike was caused by a natural trend when it was actually the result of a specific promotion. The new system, however, treats the merchant's action as a distinct, primary force. It learns a specific pattern where a past state leads to a decision, and that decision directly drives the next state. This allows the system to understand that if a merchant stops spending on ads, sales will likely drop, even if the general market trend is rising. It learns the mechanics of cause and effect rather than just the correlation of events.

The second part of the system addresses the chaos of the real world. Even with a perfect understanding of how a merchant's actions influence sales, the future is never entirely predictable. Sudden news events, viral social media trends, or unexpected holidays can cause demand to surge or crash in ways that have nothing to do with the merchant's budget. To handle this, the researchers added a correction module that acts like a fine-tuning mechanism. This module looks at external signals, such as news headlines and trending search topics, and uses advanced language processing to understand their meaning. It then calculates the difference between what the first part of the system predicted and what actually happened, attributing that difference to these external shocks. By doing this, the system can correct its own predictions, adding a layer of realism that accounts for the unpredictable nature of the world.

The researchers tested this system on a massive dataset from Alibaba's 1688 platform, which includes the sales trajectories of roughly 32 million products over a period of time. They compared their new method against several of the most powerful existing forecasting models. The results were clear and consistent. In simulations where the system was asked to predict sales under new, unseen budget plans, the traditional models struggled. They tended to stick to historical patterns, failing to react when the merchant's strategy changed. In contrast, the new system accurately followed the new path, correctly predicting how sales would rise or fall in response to the specific interventions. When the forecast horizon was extended to ten weeks, the new system reduced the error rate by more than half compared to the best-performing traditional models. It proved particularly adept at handling long-term planning, where small errors in prediction can compound into massive miscalculations.

To ensure this was not just a theoretical exercise, the team deployed the system in a live environment on the Alibaba 1688 platform. They ran a large-scale test involving hundreds of merchants who used the system to plan their advertising budgets. The merchants who used the new simulation tool saw a significant improvement in their business outcomes. On average, their return on investment increased by 15 percent, and their overall lifetime value grew by 13 percent. These gains came from the ability to allocate money more effectively, avoiding spending on campaigns that would not work and doubling down on strategies that the simulation showed would succeed. The system allowed them to see the consequences of their choices before making them, turning the planning process from a game of chance into a calculated strategy.

The success of this work suggests a shift in how we think about forecasting in complex systems. It moves the focus from simply observing the past to understanding the levers that control the future. By explicitly modeling the relationship between human decisions and market outcomes, and by separating those decisions from the noise of external events, the researchers have built a tool that is both more accurate and more useful. It demonstrates that in a world where actions constantly reshape reality, the best way to predict the future is to understand how our own choices will build it. The findings offer a new path for businesses to navigate uncertainty, providing a way to test strategies in a safe, simulated environment before committing real resources to the volatile marketplace.

Technical Summary: CEDAR

Problem Statement

In large-scale e-commerce marketplaces, time series forecasting (TSF) is critical for inventory planning, pricing, and marketing optimization. However, current forecasting approaches are predominantly passive, designed to extrapolate historical trends under observed policies. They fail to address the core need of merchants: decision-conditioned simulation. Merchants require "what-if" analysis to evaluate future sales trajectories under specific, planned intervention sequences (e.g., budget schedules, discount strategies) that may deviate significantly from historical data.

Existing multivariate TSF models typically treat merchant actions (e.g., ad spend, discounts) as auxiliary covariates concatenated with state variables. This design suffers from two main limitations:

Autoregressive Inertia: Models capture statistical correlations in the joint distribution of states and actions rather than learning explicit causal transition mechanisms. Consequently, they cannot respond to counterfactual action schedules, leading to policy-insensitive rollouts.
Conflation of Dynamics: These models entangle endogenous market evolution with decision-induced transitions and fail to separate action effects from non-stationary exogenous shocks (e.g., viral trends, holidays). This leads to unreliable counterfactual analysis and erroneous budget planning.

Methodology: CEDAR

The authors propose CEDAR (Controlled and Event-Driven Demand forecasting via Action-aware Residual decomposition), a two-stage framework designed to robustly simulate demand under planned interventions. The framework is trained on a large-scale dataset from Alibaba 1688 containing 32 million product trajectories with paired state-action sequences and aligned event signals.

Stage I: Action-Interleaved Transformer (AIT)

The first stage focuses on learning controllable action-conditioned state transitions.

Architecture: Instead of treating actions as exogenous features, CEDAR models states ( $s_t$ ) and actions ( $a_t$ ) as interleaved tokens in a unified causal sequence: $s_{t-1} \to a_t \to s_t$ .
Mechanism: An Action-Interleaved Transformer (AIT) processes this sequence. By enforcing this specific causal ordering, the model learns an explicit transition operator $s_{t+1} = f(s_t, a_{t+1})$ that captures how interventions drive subsequent state changes.
Goal: To enable stable multi-step rollouts under novel intervention strategies by disentangling controllable dynamics from passive correlations.

Stage II: Residual Correction Module

The second stage addresses exogenous disentanglement, correcting for non-stationary external factors that the AIT cannot predict from state-action history alone.

Signal Construction: The module leverages Large Language Models (LLMs) to extract structured representations from unstructured external signals. It processes news from the previous week and current holiday/seasonal events to generate "hotspot embeddings" ( $h_t$ ) that capture the semantic context of external influences.
Correction Mechanism: A cross-attention mechanism aligns these external event embeddings with item-specific temporal patterns (product metadata and recent states). A Multi-Layer Perceptron (MLP) then predicts the residual ( $\Delta s_{t+1}$ ) between the simulated trajectory and actual observations.
Final Output: The final forecast is the sum of the base prediction from Stage I and the residual correction: $\tilde{s}_{t+1} = \hat{s}_{t+1} + \Delta s_{t+1}$ .

Training Strategy

CEDAR employs a decoupled two-stage training paradigm. Stage I is trained first to learn stable transition dynamics. Stage II is trained subsequently to learn the residual corrections. This separation ensures the transition operator remains robust under policy shifts while the correction module adapts flexibly to volatile market conditions.

Key Contributions

Problem Formulation: The paper reframes merchant-facing sales forecasting as a decision-conditioned simulation problem, highlighting the inadequacy of passive TSF under sequential interventions.
Novel Framework: CEDAR introduces a two-stage, action-aware framework that:
- Learns explicit action-conditioned transitions via an Action-Interleaved Transformer.
- Corrects exogenous, non-stationary deviations using an event-driven residual module enhanced by LLM-assisted text understanding.
Empirical Validation: The approach is validated on a massive industrial dataset (32 million trajectories) and through online A/B testing in a production environment.

Experimental Results

The authors evaluated CEDAR against state-of-the-art baselines (Informer, PatchTST, PETFormer, Timer-XL) on the E-Com 15-Week dataset and a public Kaggle Store Sales dataset.

Offline Performance: CEDAR consistently outperformed all baselines in Mean Squared Error (MSE), Mean Absolute Error (MAE), and Normalized MSE (NMSE).
- In the "next 5" prediction setting, CEDAR achieved an MSE of 0.182, significantly outperforming the strongest baseline PatchTST (0.424) and PETFormer (0.434), representing relative improvements of 57.1% and 58.1%, respectively.
- In the "next 10" setting, CEDAR reduced MSE to 0.414, while baselines like Informer suffered severe degradation (MSE > 30).
Ablation Studies:
- Removing the Action-Interleaved design (using a single-stage joint training) increased MSE, confirming the necessity of explicit action-conditioned modeling.
- Removing the Residual Correction Module or shuffling temporal event signals degraded performance, particularly in capturing bursty demand fluctuations (MAE).
Online A/B Testing: Deployed on Alibaba 1688 for 30 days with 239 merchants.
- The treatment group (using CEDAR) achieved a 13% increase in Lifetime Value (LTV) (46,471 vs 41,228) and a 15% improvement in Return on Investment (ROI) compared to the control group using a diffusion-based baseline.
- The system facilitated 8.77 million RMB in transaction value with a 61% repurchase rate.

Significance and Claims

The paper claims that CEDAR provides a principled foundation for decision-aware forecasting in highly non-stationary e-commerce environments. By explicitly decoupling endogenous market dynamics from latent external shocks and treating merchant actions as first-class modalities, the framework enables reliable "what-if" analysis.

The authors emphasize that their work moves beyond passive correlation-based extrapolation to controlled simulation, allowing merchants to proactively optimize budget schedules and marketing strategies. The successful online deployment demonstrates that decision-conditioned simulation delivers tangible business value at scale, reducing ineffective ad spend and improving the alignment between budget allocation and true market demand. The paper concludes that explicit action-conditioned modeling and external shock disentanglement are essential for robust planning in intervention-driven systems.