Context

Term deposits are a key income source for banks, and telephonic marketing remains one of the most effective ways to sell them. However, these campaigns are resource-intensive, involving large call centres. To optimise efficiency, it's crucial to identify customers who are most likely to convert before reaching out. This dataset is related to direct telephonic marketing campaigns of a Portuguese bank, with the goal of predicting whether a customer will subscribe to a term deposit (yes/no). My own personal objectives are listed below.

Objectives

• Evaluate the effectiveness of the past campaign... Past Conversion Rate.
• Understand the nature of factors that actually drive conversion
• Build a propensity model to estimate the likelihood of future conversion for similar campaigns.
• Calibrate the model to better reflect probabilities
• Identify the top percentage of leads expected to convert, without triggering diminishing returns.
• Measure gains and uplift
• Design and recommend data-informed A/B experiments strategies.

Let's connect! --> LinkedIn | X | GitHub.

---

Process & Findings: Term Deposit Propensity Modelling

This analysis focused on optimising telephonic marketing campaigns for term deposit products by leveraging a robust propensity modelling approach. The goal was to identify high-likelihood conversion targets, thereby increasing campaign efficiency and ROI for the bank which led to the discovery of the fact that 83.54% of all conversions are captured by just the top 40.0% of leads, and they are 6.01 times more likely to convert than if leads are picked at random.

Key Findings

• Past Campaign Effectiveness: The historical conversion rate was 11.7%, aligning with industry norms for random targeting in direct marketing.
• Conversion Drivers:
  • Demographics: Older clients (30–60 years), those in management roles, married, and those with secondary education were more likely to subscribe.
  • Financial Status: Higher account balances correlated with higher subscription rates.
  • Behavioral Factors:
    • Longer call durations were strongly associated with successful conversions.
    • Fewer campaign contacts (1–3) were optimal; conversion rates dropped with excessive contact attempts.
    • Prior contact is kind of a tricky one, as very recent and not very recent at all, e.g 2-3 and over 100 days ago have positive outcomes. Positive outcomes from previous campaigns (poutcome) significantly boosted conversion likelihood as well.
  • Channel and Timing: Cellular contact outperformed telephone, and May was the most successful month for subscriptions.
  • Credit Status: Clients without existing loans or defaults were more receptive.

Modeling & Data Handling

• Outliers in call duration, campaign contacts, and previous contacts were carefully excluded to prevent skewed modelling.
• Categorical variables were encoded with attention to ordinal relationships (e.g., education, month).
• Data balancing techniques (SMOTE, undersampling) were considered to address class imbalance.
• The initial model was trained with a train/test/calib set to understand performance as well as the impact of calibration on score alignment.
• The final model was trained on the previously mentioned train + test set, then calibrated using the original calibrated set and sigmoid method to better reflect true conversion likelihoods.

Recommendations

1. Targeted Outreach

• Prioritise segments with high predicted propensity: older, married, highly educated, management-level clients, and those with higher balances.
• Focus on clients with no credit defaults or active loans.

2. Optimise Campaign Strategy

• Limit contact attempts to a maximum of 3 per client to avoid diminishing returns and potential customer fatigue.
• Schedule campaigns to peak in May, July, or August as they result in the highest percentage of conversions of top leads.
• Favour cellular over telephone outreach for higher engagement.

3. Personalise Messaging

• Tailor scripts and offers for high-propensity segments (e.g., highlight product benefits relevant to management professionals or retirees).
• For clients with prior positive campaign outcomes, reference past interactions to build rapport.

4. Enhance Data Quality & Monitoring

• Regularly review and update customer data to maintain segmentation accuracy.
• Try to find out the education levels and outreach methods of those currently unknown
• Monitor for outliers and update exclusion thresholds as needed to keep the model robust.

A/B Testing Recommendations

To validate and further optimise these strategies, implement the following A/B tests:

Test Focus	Group A (Control)	Group B (Test)	Success Metric
Contact Channel	Telephone outreach	Cellular outreach	Conversion rate
Campaign Timing	Standard months	May (peak month)	Conversion rate
Contact Frequency	Up to 6 contacts	Max 3 contacts	Conversion rate, opt-out rate
Personalisation	Generic script	Tailored script for high-propensity segments	Conversion rate, call duration
Prior Campaign Targeting	All clients	Only clients with prior positive outcomes	Conversion rate

• Monitor: Conversion rates, call durations, and opt-out rates for each test group.
• Iterate: Refine targeting and messaging based on statistical significance and observed lift.

Final Observations

• The propensity model provides actionable insights for resource allocation and campaign design, enabling the bank to focus efforts on clients most likely to convert.
• Regular calibration and monitoring are essential to adapt to changing customer behaviours and market conditions.
• Combining data-driven targeting with thoughtful A/B testing will maximise marketing ROI and customer satisfaction.

NOTE: This entire analysis is part of my weekly series in efforts to demystify applied statistical techniques through real-world, project-driven examples, making concepts like propensity modelling, causal inference, and evaluation metrics more accessible to practitioners of all backgrounds.