Predicting Costco Member
Holiday Spending

Using Simulated Data to Compare Statistical and Machine Learning Models

Zeina Abudamous

Business Problem

Project Goals

  • Predict how much a Costco member will spend during the Holiday season (from October to December) based on their past behaviors.

  • Asses member behavior to see which characteristics predict holiday spend best in order to support targeted promotions and inventory planning.

Data Design

Data Overview

  • 5,000 Simulated Costco members (rows)

  • 14 behavioral and membership features (columns)

Why Simulated Data?

  • Costco has a very strict privacy policy, and does not share member data.

  • Most simulated data takes random numbers from a Normal distribution, I wanted more control over the data design.

Feature Engineering

  • Member behavior was simulated using appropriate statistical distributions.

    • Tenure modeled with a right-skewed beta distribution to reflect high renewal rates.

    • Spending generated with gamma distributions to capture skewed, positive-only purchase behavior.

    • Visit counts modeled using Poisson processes (event-based behavior).

    • Department-level spending shares generated via a Dirichlet distribution to ensure proportions sum to 1.

Target Variable

  • Constructed as a nonlinear function of

    • Prior spending, category mix, membership status (executive vs. gold star)

  • It follows a right-skewed distribution, consistent with typical retail data.

  • Applied log transformation to reduce influence of extreme spenders.

Feature Correlation Analysis

  • Highly correlated features:
    rolling_12mo_spend & rolling_12mo_avg_basket: 0.83

  • Features retained for XGBoost, but rolling_12mo_avg_basket was removed for Multiple Linear Regression and ElasticNet

Model Performance

  • The tree based model (XGBoost) performed slightly better than the linear models (Linear Regression & ElasticNet).
Model \(R^2\) MAE RMSE
Linear Regression 0.6614 $264.66 $333.65
ElasticNet 0.6625 $264.42 $333.11
XGBoost 0.6864 $256.20 $321.07

Individual Prediction Analysis

  • When the actual Holiday Spend is $1,460.97 Each model’s prediction is:
Model Prediction Error
Linear Regression $1,158.06 $302.91
ElasticNet $1,151.33 $309.63
XGBoost $1,190.66 $270.31

Feature Importance

  • Looking at the SHAP results of the directional effects for individual predictions, prior annual spending dominates predictions.

Business Takeaways

Key Insights:

  • Spending history is the strongest predictor

  • Executive membership status also impact spending

Recommendations:

  • Use XGBoost for holiday spend predictions.

  • Focus on members with high spend history for holiday campaigning.