BANK MARKETING STUDY · 2026
Term Deposit
Subscription
Prediction
Comparing Logistic Regression and Random Forest to identify customers most likely to respond to a direct-marketing campaign.
R 4.5.2caretrandomForestpROCPCAVIF
RESEARCH OVERVIEW
Who is likely to subscribe?
The study predicts whether a customer will subscribe to a bank term deposit using demographic, financial, contact, economic, and previous-campaign information. It contrasts an interpretable parametric model with a flexible ensemble learner.
RQ
Research questionWhich model best identifies likely subscribers while producing useful evidence for campaign decisions?
01 · DATASET
Prepared for realistic pre-call prediction.
Target distribution
Most customers did not subscribe.
Accuracy alone is misleading when 88.73% belong to the “no” class, so recall, precision, F1, and ROC-AUC were essential.02 · EXPLORATORY ANALYSIS
Campaign response is shaped by several signals.
Duration removed
LeakageCall duration is only known after contact and would create unrealistic performance.
Repeated contact skews
RightMost customers were contacted only a few times, suggesting diminishing returns.
Economic overlap
High VIFEuribor, employment variation, and employee count showed multicollinearity.
Compact structure
70%The first three principal components explained about 70% of numeric variance.
Five predictor families
DemographicAge · job · education
FinancialDefault · housing · loan
CampaignContact · month · attempts
Important campaign signals
Previous outcomepoutcome
Economic climateemp_var_rate
Timing and channelmonth · contact
03 · METHODOLOGY
A leakage-safe modelling pipeline.
01CleanStandardise names; retain unknown levels
→02TransformFactors and Z-score numeric scaling
→03Split70% train · 30% test · stratified
→04ModelTwo logistic and two forest variants
→05EvaluateAccuracy, precision, recall, F1, AUC
Logistic Regression
Coefficients and odds ratios translate directly into campaign guidance, with VIF checks supporting responsible interpretation.
Random Forest
Tree ensembles model interactions and rank variable importance without relying on normal distributions.
Recall matters
Finding genuine subscribers has greater campaign value than simply predicting the majority class correctly.
04 · MODEL PERFORMANCE
Interpretability versus minority-class detection.
ModelAccuracyPrecisionRecallF1AUCTreesBest use
Logistic Regression 1 BEST AUC.9004.6695.2284.3407.7888—Interpretation
Logistic Regression 2.8996.6624.2227.3333.7884—Simplified
Random Forest 1.8971.5872.2902.3885.7783300Recall
Random Forest 2 BEST F1.8979.5948.2931.3927.7799500Targeting
05 · INTERPRETATION
Previous outcomes and economic context matter.
Previous outcome100
Employment rate86
Consumer prices78
Contact month72
Contact channel64
Euribor 3m59
Campaign contacts48
BUSINESS FINDING
Model choice depends on campaign goals.
Logistic Regression offered the strongest AUC and clearer managerial explanation. The tuned Random Forest found more actual subscribers and achieved the best F1 score, making it more useful where positive-case identification is the priority.
High accuracy does not guarantee an effective marketing model when the valuable response class is rare.
06 · CONCLUSION
Use interpretable probability and class-aware targeting together.
Logistic Regression achieved the best overall discrimination at AUC 0.7888, while the 500-tree Random Forest improved recall to 0.2931 and F1 to 0.3927.
The analysis demonstrates why leakage prevention, stratified evaluation, and business-relevant metrics are essential when building real-world campaign models.