[Paper Review] A multi-series framework for demand forecasts in E-commerce

Rémy Garnier, Arnaud Bellétoile|arXiv (Cornell University)|May 31, 2019

Forecasting Techniques and ApplicationsDecision Sciences12 references3 citations

TL;DR

This paper proposes a global XGBoost-based framework for e-commerce demand forecasting that leverages cross-series information and non-linear relationships to improve accuracy on short, volatile sales time series. By incorporating hierarchical product groupings, seasonality features, and advanced preprocessing to handle 'fake zeros' and sparse data, the model reduces RMSE by 10% and MAE by 5% compared to state-of-the-art benchmarks, with significant gains in early product lifecycle forecasting.

ABSTRACT

Sales forecasts are crucial for the E-commerce business. State-of-the-art techniques typically apply only univariate methods to make prediction for each series independently. However, due to the short nature of sales times series in E-commerce, univariate methods don't apply well. In this article, we propose a global model which outperforms state-of-the-art models on real dataset. It is achieved by using Tree Boosting Methods that exploit non-linearity and cross-series information. We also proposed a preprocessing framework to overcome the inherent difficulties in the E-commerce data. In particular, we use different schemes to limit the impact of the volatility of the data.

Motivation & Objective

• To address the challenge of short, volatile sales time series in e-commerce, where univariate models fail due to limited historical data.
• To improve forecasting accuracy by sharing information across related product series through a global modeling approach.
• To enable cold-start predictions for new products with minimal historical data by exploiting group-level patterns and covariates.
• To overcome data-specific challenges such as 'fake zeros', non-stationarity, and bursty sales through a tailored preprocessing pipeline.
• To evaluate the framework on real-world e-commerce data and demonstrate superior performance over industry benchmarks and state-of-the-art methods.

Proposed method

• The framework uses XGBoost to model demand forecasts globally across all products, learning from both individual and cross-series patterns.
• It incorporates three types of covariates: temporal (e.g., holidays), longitudinal (e.g., product category), and mixed (e.g., weekly prices) to enrich feature representation.
• Seasonality features are derived from aggregated behavior of similar products, enabling robust estimation even with short individual series.
• A preprocessing pipeline identifies and imputes 'fake zeros' (e.g., due to stockouts) using univariate forecasting on non-zero sales segments.
• Categorical features are encoded via ordinal or hashing encoding, with ordinal encoding showing superior performance in experiments.
• The model is trained end-to-end on a large-scale e-commerce dataset with hierarchical product groupings and temporal features, using RMSE and MAE as evaluation metrics.

Experimental results

Research questions

• RQ1 Can a global boosting model that shares information across product series outperform univariate forecasting methods on short, volatile e-commerce sales time series?
• RQ2 How effective is the use of hierarchical product groupings and cross-series information in improving forecast accuracy for products with limited historical data?
• RQ3 To what extent do seasonality features derived from group-level behavior enhance forecasting performance on individual short series?
• RQ4 Does the proposed preprocessing pipeline effectively mitigate the impact of 'fake zeros' and data sparsity in e-commerce demand forecasting?
• RQ5 How does the model perform during the early lifecycle of a product, where cold-start prediction is most critical?

Key findings

• The XGBoost-based global model reduces RMSE by 10% and MAE by 5% compared to the industry benchmark across the full dataset.
• The model significantly outperforms the benchmark in early product lifecycle stages, reducing RMSE by 42.5% and MAPE by 24.0% for products with only 10 weeks of history.
• Ordinal encoding of categorical features consistently outperformed hashing encoding, suggesting better preservation of feature structure.
• Incorporating seasonality features improved performance, especially for high-volume products in group A, which have the highest business impact.
• The model enables effective cold-start forecasting for new products, achieving strong performance even with minimal historical data.
• The framework demonstrates robustness to data volatility and sparsity, with consistent gains across all product categories (A, B, and C).