Term Deposit Marketing Prediction & Subscriber Segmentation

Project Overview

At Apziva, I designed and delivered a machine-learning solution for a European bank’s term-deposit campaign. Our goals were to:

1. Predict which customers would subscribe to a term deposit
2. Segment the subscriber base into actionable sub-groups

This enabled the client to focus on high-value prospects and tailor their outreach scripts and offers.

---

Project Workflow

1. Data Ingestion & Definition

   • Loaded 40 000 call records with 14 attributes (age, job, balance, contact details, call metadata, etc.)
   • Defined numeric vs. categorical fields for consistent processing

2. Exploratory Data Analysis (EDA)

   • Uncovered the strongest conversion drivers
   • Benchmarked subscriber vs. non-subscriber behavior

3. Train/Test Split

   • Performed a 75/25 stratified split (seed=42)

4. Model Selection & Tuning

   • Used PyCaret to compare models and tune hyperparameters
   • Selected LightGBM and further tuned with PyCaret

5. Feature Selection

   • Identified the top predictors for conversion

6. Subscriber Extraction (SQL)

   • Isolated the 2 896 subscribers (7.24% conversion) via DuckDB

7. Subscriber Segmentation (K-Means)

   • Ran K-Means (k=4) on scaled numeric features to discover four distinct subscriber personas

8. Insights & Business Impact

   • Synthesized findings into targeted recommendations

---

EDA Key Insights

• Call Duration

  • Median for subscribers: 629 s vs. 164 s for non-subscribers
  • Longer conversations drive higher conversions

• Account Balance

  • Median balance: €620 (subscribers) vs. €395 (non-subscribers)
  • Wealthier customers are more likely to subscribe

• Campaign Attempts

  • Mean attempts before conversion: 2.41 (subscribers) vs. 2.92 (non-subscribers)
  • Diminishing returns after two calls

• Age

  • Median age: 37 (subscribers) vs. 39 (non-subscribers)
  • Only a small difference, other factors are more decisive

• Job Segments

  • Highest conversion rates:
    • Students: 15.65 %
    • Retirees: 10.51 %
    • Unemployed: 8.70 %
  • Lowest conversion rates:
    • Housemaids: 4.88 %
    • Blue-collar: 5.70 %

---

Modeling & Feature Importance

After comparing multiple algorithms, LightGBM delivered the best balance of precision and recall on the "yes" class.

Top 5 Predictive Features (Random Forest importances):

Feature	Relative Importance
duration	37.7%
balance	12.0%
age	10.4%
day	10.1%
campaign	4.6%

---

Subscriber Segmentation (k = 4)

Clustered the 2896 subscribers using numeric features (age, balance, duration, campaign, day). Four meaningful segments emerged:

Cluster	Size	Avg Age	Avg Balance	Avg Duration	Avg Attempts	Characteristics
0	774	53.2	€1,429	586 s	1.92	Older, low balance, moderate call lengths
1	1419	32.8	€1,047	529 s	1.77	Young adults, moderate balance & durations
2	116	42.1	€12,217	627 s	2.42	Mid-age, very high balances
3	587	38.8	€1,009	1197 s	4.64	Moderate age/balance, very long & repeated calls

Actionable Takeaways:

• Cluster 2 (High-Balance subscribers): Prioritize for premium term-deposit offers.
• Cluster 3 (Engaged callers with long durations): Tailor detailed product walkthroughs as they respond to extended conversations.
• Cluster 1 (Younger customers): Offer entry-level or flexible-term products.
• Cluster 0 (Older, low-balance): Bundle marketing with cross-sell opportunities (e.g., savings or retirement planning).

---

Conclusion & Next Steps

• Business Impact:
  • Focusing on high-value segments could lift overall campaign ROI by an estimated 15 %.
  • Tailored scripts per cluster enable more efficient agent time allocation and stronger customer engagement.

I welcome any questions and look forward to discussing how these techniques can be applied to your data-driven marketing challenges.