Forecasting own brand sales: Does incorporating competition help?

Wei Li, Dennis Fok, Philip Hans FransesRePub 2019

TL;DR

This paper investigates whether including competitor information improves sales forecasts for own brands. The study finds that modern machine learning techniques like Lasso, Elastic Net, and Random Forests significantly outperform traditional benchmark models by incorporating competitive data, with shrinkage methods offering a safe and effective approach for most brands.

What Problem Does This Paper Address?

Accurate sales forecasting is crucial for retailers and manufacturers to manage inventory, marketing, and production. Traditional sales forecasting models often rely solely on a brand's own historical sales and marketing activities. However, in competitive markets, understanding and incorporating the actions of rival brands can potentially lead to more accurate predictions. The challenge lies in effectively integrating this competitive information, especially when the set of relevant competitors is large and their impact is complex and non-linear. This paper aims to quantify the value added by incorporating competitive information and to compare various modern forecasting techniques for this purpose.

Methods

The study compares several forecasting methods:

  • Benchmark Model: An autoregressive model using only the focal brand's own sales and marketing data.
  • Average Competitor Model (ACM): Summarizes competitor information by taking market-share weighted averages of their sales and marketing variables.
  • Principal Component Regression (PCR): Uses Principal Component Analysis (PCA) to reduce the dimensionality of competitor data into a smaller set of factors, which are then used in a regression model. This includes variations like "Grouped PCR" which applies PCA to specific groups of variables (price, feature, display, sales).
  • Hard and Soft Thresholding: These methods select a subset of relevant competitor variables based on their statistical significance or correlation with the target variable, and then construct factors from these selected variables for forecasting.
  • Shrinkage Methods (Lasso and Elastic Net): These techniques fit models with all predictor variables but apply penalties to shrink coefficients towards zero. Lasso specifically forces coefficients of less important predictors to exactly zero, performing variable selection. Elastic Net combines Lasso and Ridge regression penalties.
  • Tree-Based Models (Random Forest and Boosting): These methods build multiple decision trees and aggregate their predictions. Random Forest randomly selects predictor subsets at each split, while Boosting sequentially builds trees, with each new tree correcting the errors of the previous ones.

The study also explores two forecasting window approaches: expanding window (increasing the estimation window with new data) and rolling window (fixed-size window sliding over time).

Datasets

The study utilizes weekly sales data for 169 brands across 31 product categories, sourced from Information Resources Inc. (IRI). This dataset spans over five years and includes information on brands' own sales, marketing activities (price, display, feature variables), and competitor data. The data was aggregated to the brand level by weights, and stock-keeping units (SKUs) were grouped into narrower categories.

Key Insights and Ideas

  • Competition Matters: Incorporating competitor information significantly improves sales forecasts for the vast majority of brands, with the benchmark model being outperformed for 162 out of 169 brands.
  • Shrinkage Methods as a Safe Bet: Lasso and Elastic Net consistently offer improved forecast accuracy over the benchmark for most brands and are considered "safe" choices due to their robust performance.
  • Random Forest for Niche Performance: While not always outperforming the benchmark for every brand, Random Forest can achieve substantial accuracy improvements for specific brands, particularly those with infrequent promotions where other models struggle.
  • Data-Driven Factor Selection is Key: For factor-based models (PCR, Hard/Soft Thresholding), using a data-driven approach to select the optimal number of factors consistently yields better results than fixing the number of factors.
  • Tree Models Handle Non-Linearity: Tree-based methods like Random Forest are adept at capturing complex non-linear relationships and interactions, which linear models often miss, explaining their strong performance in certain scenarios.

Key Results

  • Overall Improvement: The study found that 162 out of 169 brands showed improved forecast accuracy when competitor information was incorporated.
  • Best Performing Models: Random Forest achieved the best performance in terms of the lowest mean and median MAPE (17.0% and 21.5% respectively) and the highest number of brands with MAPE below 20% and 30%. Elastic Net and Lasso also showed strong results, with Elastic Net leading in terms of brands performing better than the benchmark for RMSE (117 brands) and MAE (130 brands).
  • Benchmark Performance: The benchmark model (without competitor data) had a median MAPE of 18.9% and a mean MAPE of 24.1%.
  • Computational Cost: Shrinkage methods (Lasso, Elastic Net) had a comparable computation burden to factor models, while Random Forest was significantly more computationally intensive. However, the paper notes that parallel computing can mitigate this.
  • Category-Specific Performance: Analysis by product category revealed that Random Forest performed best in many categories, particularly in cigarettes, mayonnaise, and razors, while Lasso and Elastic Net showed broader, consistent improvements.

Limitations and Future Work

  • Variable Variation Threshold: The study excluded variables with fewer than five distinct values, which might have led to the exclusion of valuable promotion information for products with infrequent promotions. The authors suggest that tree models are less sensitive to this limitation.
  • Cross-Validation for Time Series: The paper acknowledges that using standard cross-validation for time series data is not strictly correct due to interrupted temporal dependencies, but it is used as a heuristic.
  • Future Research Directions:
    • Investigating ways for shrinkage methods to better incorporate infrequent promotion information, possibly by setting upper bounds on forecasted sales.
  • Forming a "committee" of the best-performing models (Lasso, Elastic Net, Random Forest) to leverage their respective strengths.

Summary updated

Abstract

This study aims to investigate how much value is added to traditional sales forecast-
\ning models in marketing by using modern techniques like factor models, Lasso, elastic
\nnet, random forest and boosting methods. A benchmark model uses only the focal
\nbrand's own information, while the other models include competitive sales and market-
\ning activities in various ways. An Average Competitor Model (ACM) summarises all
\ncompetitive information by averages. Factor-augmented models incorporate all or some
\ncompetitive information by means of common factors. Lasso and elastic net models
\nshrink the coecient estimates of specic competing brands towards zero by adding a
\nshrinkage penalty to the sum of squared residuals. Random forest averages many tree
\nmodels obtained from bootstrapped samples. Boosting trees grow many small trees
\nsequentially and then average over all the tree models to deliver forecasts. We use these
\nmethods to forecast sales of packaged goods one week ahead and compare their pre-
\ndictive performance. Our
BibTeX
@inproceedings{li2019forecasting,
  author = {Wei Li and Dennis Fok and Philip Hans Franses},
  title = {{Forecasting own brand sales: Does incorporating competition help?}},
  booktitle = {RePub},
  year = {2019},
}
Figure