Agenda

• Background & data description
  
• Key exploratory findings
  
• Data preparation & transformations
  
• Candidate models (full vs. reduced)
  
• Model selection (VIF + cross-validation)
  
• Final model results & interpretation
  
• Test-set performance
  
• Conclusions & limitations

Background

• Customer information for a clothing store in New England

• Includes sales, visit frequency, product mix, and promotion history

• Goal: predict response to a direct-mail promotion to optimize targeting

Description of Data Set

• Source: clothing store data set from a regional retail chain
• Data pulled from Kaggle
• Observations: 2,619 customers (after removing missing values)
• Variables: 20 customer attributes, including:
  • Purchase behavior: frequency, spending, product mix
  • Promotion history: number of promos received, past engagement
  • Account characteristics: days on file, returns, cluster type
• Response Variable: whether a customer responded to a direct-mail offer (yes/no)

Practical Questions

• Who is most likely to respond to the promotion?
  • Can we identify high-engagement customer profiles?
• Which customer behaviors drive response likelihood?
  • Does shopping frequency matter more than dollar spending?
• Does promotion saturation affect engagement?
  • Do customers respond less when they receive too many offers?
• Can we create a practical model for targeting?

Key Data Insights

Exploratory Findings

• Customer purchase & spending features are strongly right-skewed

• Promotional history varies widely

• Several purchase metrics move together → risk of multicollinearity

Model Implications

• Use log transformations on skewed purchasing variables

• Check redundancy among correlated predictors

• Model choice: Logistic Regression for response prediction

Key Data Insights (Continued)

Exploratory Data Analysis

• Purchase & spending metrics move together (e.g., FRE, MON, STYLES)

• More visits ⇒ more items ⇒ more total sales

• Right-skewed distributions → transformations before modeling

Data Preparation for Modeling

Train/Test Split

• 70% of customers used to train the model

• 30% held out for testing (not touched during training)

• Evaluates real predictive performance on unseen customers

• All transformations were applied only to the training data

Candidate 1: Full Model

VIF Results for Candidate 1

##       MON_log       FRE_log    STYLES_log   CLASSES_log    STORES_log 
##         12.04          5.52         15.81          9.95          1.52 
##        HI_log PERCRET_logit        PROMOS          DAYS           GMP 
##          3.56          2.07          2.27          1.86          1.70

• Strong multicollinearity among purchase variables
• Model is over explaining the same behavior

Reducing Redundancy

• We want to keep a simpler set of predictors while still representing:
  • purchasing engagement
  • advertisement exposure
  • product variety
• Let’s retain:
  • number of purchase visits (FRE_log)
  • number of items purchased (STYLES_log)
  • number of marketing promos on file (PROMOS)

Candidate 2: Reduced Model

VIF Results for Reduced Model

##    FRE_log STYLES_log     PROMOS 
##       4.29       4.29       1.66

Cross Validation for Model Selection

• 5-fold CV on training data only
• Aims to estimate prediction error on unseen customers
• Compare Candidate 1 vs Candidate 2
• Select model with lowest classification error

Model Selection

Reduced Logistic Regression Model

Why?

• Reduced model slightly outperformed full model

• Reduced model gets rid of redundant variables

• Lower multicollinearity and easier to interpret

Results

FRE_log (purchase frequency)
- Higher shopping frequency strongly increases the odds of responding
- Customers who shop more often are much more likely to engage with promotions

STYLES_log (product variety)
- Buying a wider variety of items is associated with higher response odds
- Variety shoppers seem more engaged with the brand overall

PROMOS (number of promotions on file) - - Odds ratio < 1 → each additional promo slightly reduces response odds
- Suggests promotion fatigue

Test Set Performance

• Model correctly classifies 660 out of 786 customers
• 88.3% specificity (very strong at predicting non-responders)
• 44.2% sensitivity (moderate at detecting responders)

Conclusions

• Customer behavioral features can effectively predict promotion response

• More frequent shoppers and those buying more styles are much more likely to respond

• Customers with many promotions on file are less likely to respond → evidence of promotion fatigue

• Reduced model:

  • Uses only 3 variables (FRE_log, STYLES_log, PROMOS)
  • Matches or slightly improves accuracy vs. full model
  • Is easier for a marketing team to understand and deploy

Limitations & Future Work

• Model is based on data from a single retailer → may not generalize to other regions or brands

• Only uses a subset of available variables:

  • No demographic data
  • No channel preference (e.g., web vs. in-store behavior)

• Future improvements:

  • Test nonlinear models (e.g., random forests, gradient boosting)
  • Explore different probability thresholds for specific campaign goals