[Paper Review] Improved Sales Forecasting using Trend and Seasonality Decomposition with LightGBM

Tong Zhou|arXiv (Cornell University)|May 26, 2023

Time Series Analysis and Forecasting4 citations

TL;DR

This paper proposes a hybrid sales forecasting framework that decomposes time series into trend and seasonality components, groups products based on these components, and applies LightGBM with a Tweedie loss function for improved accuracy. The method achieves a 61.4% reduction in WRMSSE compared to baseline models, demonstrating superior performance on Walmart sales data.

ABSTRACT

Retail sales forecasting presents a significant challenge for large retailers such as Walmart and Amazon, due to the vast assortment of products, geographical location heterogeneity, seasonality, and external factors including weather, local economic conditions, and geopolitical events. Various methods have been employed to tackle this challenge, including traditional time series models, machine learning models, and neural network mechanisms, but the difficulty persists. Categorizing data into relevant groups has been shown to improve sales forecast accuracy as time series from different categories may exhibit distinct patterns. In this paper, we propose a new measure to indicate the unique impacts of the trend and seasonality components on a time series and suggest grouping time series based on this measure. We apply this approach to Walmart sales data from 01/29/2011 to 05/22/2016 and generate sales forecasts from 05/23/2016 to 06/19/2016. Our experiments show that the proposed strategy can achieve improved accuracy. Furthermore, we present a robust pipeline for conducting retail sales forecasting.

Motivation & Objective

• To address the challenge of accurate retail sales forecasting amid complex trends, seasonality, and data heterogeneity across products and locations.
• To improve forecast accuracy by grouping time series based on unique trend and seasonality patterns rather than treating all products uniformly.
• To evaluate the effectiveness of a Tweedie-based loss function in handling zero-inflated, skewed sales data common in retail.
• To develop a scalable, robust pipeline for large-scale sales forecasting that integrates decomposition, model-specific learning, and ensemble weighting.

Proposed method

• The authors propose a novel measure to quantify the unique impact of trend and seasonality on each time series, enabling data-driven grouping of products with similar behavior.
• Time series are decomposed using the Prophet model to extract trend and seasonal components, which are then used to compute group-specific scores for clustering.
• LightGBM is trained on each group using features derived from the decomposed components, with a Tweedie loss function to handle zero-inflated and highly variable sales data.
• A non-recursive forecasting strategy is employed to prevent error propagation across the 28-day prediction horizon.
• Final forecasts are generated by combining LightGBM and Prophet predictions using weighted averaging, where weights are derived from LightGBM’s attention scores.
• The framework uses a hybrid modeling approach: LightGBM captures complex non-linear patterns within groups, while Prophet ensures robustness to seasonality and irregularities.

Figure 1: Total sales per day

Experimental results

Research questions

• RQ1 Can grouping time series based on trend and seasonality decomposition improve sales forecasting accuracy compared to global modeling?
• RQ2 Does the use of a Tweedie loss function enhance model performance on zero-inflated, skewed sales data?
• RQ3 How does the proposed decomposition-based grouping strategy compare to models without decomposition or with alternative loss functions?
• RQ4 To what extent does non-recursive forecasting mitigate error accumulation in multi-horizon sales predictions?

Key findings

• The proposed trend-seasonality decomposition strategy significantly improves forecast accuracy, reducing WRMSSE from 0.963 (no decomposition, MSE) to 0.767 (MSE) and to 0.614 (Tweedie loss).
• The model with both decomposition and Tweedie loss achieved the lowest WRMSSE of 0.614, outperforming linear regression (1.120), SVM (1.083), and LSTM (0.802).
• The Tweedie-based loss function consistently outperformed mean squared error (MSE) across all configurations, indicating its suitability for skewed, zero-inflated sales data.
• The non-recursive forecasting approach effectively prevented error accumulation, ensuring stable performance across the 28-day prediction window.
• Grouping time series based on trend and seasonality components led to more accurate and tailored forecasts than treating all products uniformly.
• The hybrid framework combining LightGBM and Prophet demonstrated robustness and scalability, making it suitable for large-scale retail forecasting applications.

Figure 2: Sales trends in 10 stores