💬 NLP

EventCast: Hybrid Demand Forecasting in E-Commerce with LLM-Based Event Knowledge

EventCast is a modular forecasting framework that improves e-commerce demand prediction during high-impact periods by using Large Language Models to transform unstructured business data into interpretable event summaries, which are then fused with historical time-series data via a dual-tower architecture.

Original authors: Congcong Hu, Yuang Shi, Fan Huang, Yang Xiang, Zhou Ye, Ming Jin, Shiyu Wang

Published 2026-02-12

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Congcong Hu, Yuang Shi, Fan Huang, Yang Xiang, Zhou Ye, Ming Jin, Shiyu 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

Imagine you are running a small bakery. Most days, you know exactly how many croissants to bake because you look at what you sold last Tuesday. This is "Historical Forecasting"—relying on the past to predict the future.

But then, something unexpected happens. Suddenly, there’s a massive local festival, or a sudden heatwave, or a celebrity posts a video eating your sourdough. If you only look at last Tuesday’s numbers, you’ll either bake way too little (and run out) or way too much (and waste money).

The researchers at ByteDance created EventCast to solve this exact problem for giant e-commerce companies. Here is how it works, explained through a few simple analogies.

1. The Problem: The "Blindfolded Chef"

Most current AI forecasting models are like a chef who is blindfolded. They can feel the temperature of the oven and remember how many loaves they baked yesterday, but they have no idea what is happening outside the kitchen. They don't know it’s Ramadan, they don't know there’s a "Flash Sale" happening, and they don't know that a holiday is about to shut down all the delivery trucks. Because they are "blind" to these upcoming events, they get caught completely off guard when demand suddenly spikes or crashes.

2. The Solution: The "Smart Assistant" (The LLM)

Instead of trying to make the "Chef" (the math model) smarter at math, the researchers gave the Chef a Smart Assistant—a Large Language Model (like ChatGPT).

However, they didn't ask the Assistant to do the cooking (the actual number crunching). Why? Because LLMs are great at talking, but they are actually quite bad at precise math. If you ask an LLM, "Exactly how many units of shampoo will we sell at 2:14 PM next Thursday?" it might give you a very confident, but totally wrong, number.

Instead, the Assistant's job is Reasoning. The Assistant reads the "messy notes" from the business team—notes that might have typos, slang, or weird abbreviations like "B2G1 free!! FS for orders >15USD"—and translates them into a clear summary:

"Hey Chef, next Thursday is a major holiday. People will be shopping heavily in the morning, but delivery services will be closed in the afternoon. Also, there is a 'Buy One Get One' sale active."

3. The Architecture: The "Dual-Tower" Strategy

EventCast uses what they call a "Dual-Tower" design. Think of it like a professional sports team:

Tower 1 (The Veteran Player): This part of the model looks at the "stats." It studies the historical patterns, the trends, and the seasonal cycles. It’s the expert on what usually happens.
Tower 2 (The Scout): This part listens to the Smart Assistant. It focuses entirely on the "upcoming news"—the festivals, the sales, and the holidays.

The Fusion: At the end, the two towers meet. The model takes the "usual trend" from the Veteran and adjusts it based on the "news" from the Scout. If the Veteran says, "We usually sell 100 items," but the Scout says, "There's a massive 50% off sale tomorrow," the model combines those signals to predict a surge.

4. Why is this a big deal? (The Results)

In the real world (across 4 countries and 160 regions), this worked incredibly well.

When "crazy" events happened—like huge holiday sales or sudden promotions—the old models failed miserably. But EventCast was able to "see" the event coming. It improved accuracy by massive margins (up to 86% to 97% better than models without event knowledge).

Summary in a Nutshell

EventCast is like giving a math-heavy forecasting machine a pair of "semantic glasses." It doesn't change how the machine does math; it just allows the machine to read the news so it isn't surprised when the world changes.

Technical Summary: EventCast: Hybrid Demand Forecasting in E-Commerce with LLM-Based Event Knowledge

EventCast is a modular, hybrid forecasting framework designed to address the volatility of e-commerce demand caused by non-stationary events such as flash sales, religious holidays, and sudden promotional interventions.

1. The Problem: The "Event-Driven" Forecasting Gap

Traditional demand forecasting models (both statistical and deep learning-based) primarily rely on the "self-stimulation assumption," which assumes future patterns are extensions of historical trends. This approach fails in e-commerce due to three specific challenges:

Non-stationary Demand: High-impact events (e.g., Ramadan, Black Friday) create abrupt surges or drops that historical data alone cannot predict.
Unstructured Knowledge: Critical business information (campaign details, seller incentives, logistics changes) is often stored in unstructured, "noisy" text (typos, abbreviations, inconsistent formats) that is difficult for traditional models to parse via rule-based feature engineering.
The LLM Paradox: While Large Language Models (LLMs) possess vast world knowledge, they are poor direct numerical forecasters due to low precision, high computational costs, and poor temporal consistency.

2. Methodology: Decoupling Reasoning from Prediction

EventCast introduces a dual-tower architecture that separates semantic reasoning from numerical forecasting. Instead of using the LLM to predict numbers, it uses the LLM as a domain-aware reasoning agent.

The Pipeline consists of four stages:

Event Expert Database: The system leverages existing, unstructured operational databases (campaign calendars, holiday schedules, and informal business reports) containing "noisy" text.
LLM-Based Semantic Reasoning: A frozen LLM receives a parameterized prompt containing the target date and the relevant database snippets. The LLM performs step-by-step reasoning to interpret cultural nuances (e.g., the difference between shopping surges and religious observance) and complex promotion mechanics (e.g., "buy 2 get 1 free" or "stackable coupons"). It outputs a structured, interpretable textual summary within <result> tags.
Feature Embedding and Alignment:
- Historical Tower: A Transformer encoder processes multivariate historical time-series features.
- Event Tower: The LLM's textual summaries are tokenized into learnable embeddings and aggregated into a semantic vector.
- Fusion: The two towers are aligned in a shared latent space using a linear mapping and fused via a residual-connected Feed-Forward Network (FFN).
Dual-Head Forecasting: The model splits the prediction into two interpretable components:
- $\hat{y}_{trend}$ : The baseline demand predicted from historical patterns.
- $\hat{y}_{event}$ : The adjustment factor driven by future event signals.
- The final forecast is a weighted combination: $\hat{y} = \lambda \cdot \hat{y}_{trend} + (1 - \lambda) \cdot \hat{y}_{event}$ .

3. Key Contributions

Hybrid Architecture: Integrates high-dimensional historical signals with low-dimensional, high-semantic event summaries.
Efficiency & Scalability: By using only the textual output of the LLM rather than its heavy embeddings or fine-tuning, the model remains lightweight, easy to maintain, and allows for "plug-and-play" updates (e.g., swapping for a better LLM).
Interpretability: Unlike "black-box" models, EventCast provides human-readable reasoning traces, allowing warehouse and category managers to understand why a forecast has changed.
Robustness to Noise: The LLM naturally handles typos, abbreviations, and complex business logic that would break traditional string-matching or rule-based systems.

4. Results and Empirical Evidence

The framework was deployed across 4 countries and 160 regions over 10 months.

Overall Performance: EventCast outperformed state-of-the-art baselines (like ChronosX and iTransformer), achieving significant reductions in MAE and MSE.
Event-Driven Superiority: The most dramatic gains occurred during high-volatility periods. Compared to the best industrial baseline, EventCast reduced MAE by up to 57.0% and MSE by 83.3% during event-driven windows.
LLM Utility: The study proved that using LLMs for symbolic reasoning (extracting context) is far superior to using them for direct forecasting (predicting numbers), as the latter suffers from quantization errors and autoregressive drift.
Ablation Insights: Removing the future event knowledge significantly degraded performance, especially at longer forecasting horizons ( $T+4$ ), where the value of "knowing the future" is highest.

5. Significance

EventCast represents a paradigm shift in industrial time-series forecasting. It demonstrates that the most effective way to utilize LLMs in production is not to replace specialized numerical models, but to use them as semantic bridges that translate unstructured human business intelligence into structured signals that mathematical models can consume. This makes the system highly practical for real-world, fast-moving e-commerce environments.