In response to declining performance across multiple telemarketing campaigns, a leading Portuguese bank sought a more effective strategy to cross-sell time deposit products. With an average conversion rate of only 11.7%, the challenge was to identify which customers were most likely to respond positively—so that future campaigns could be smarter, more efficient, and far more profitable.

I led the analysis to:

• Predict which customers are likely to subscribe
• Understand the key behavioral and demographic drivers of subscription
• Design a data-backed campaign strategy that reduces costs, improves conversion rates, and builds long-term customer value

The bank's main issue was poor targeting—telemarketers were calling almost every customer, without clarity on who was most likely to say yes. Each call cost approximately $2, including phone and representative time (~4.3 minutes per customer). This inefficient approach diluted ROI and strained marketing budgets.

To answer these questions, I followed a complete CRISP-DM (Cross-Industry Standard Process for Data Mining) approach, combining business thinking with analytical rigor.

1. Business Understanding

My goal was to reverse the bank's inefficient approach by identifying high-probability subscribers and creating a cost-optimized marketing strategy that would increase ROI and reduce wasteful spending.

2. Data Understanding

The dataset included 45,211 records from past campaigns, each representing one customer contact. Key data fields included:

• Customer demographics (age, job, marital status, education)
• Financial status (balance, loan status, housing)
• Interaction data (contact method, call duration, month)
• Past marketing performance (previous outcome, number of contacts)

This holistic dataset gave us a full picture of both who the customer is and how they respond to marketing.

3. Data Preparation

I performed:

• Variable transformation and encoding for categorical data
• Training-validation-prediction splits (60–30–10)
• Feature selection using forward stepwise logistic regression

4. Predictive Modeling

I selected logistic regression as the primary model due to:

• Its interpretability (key for non-technical stakeholders)
• Strong predictive performance
• Easy scalability and transparency in how features affect subscription odds

From the model, I identified the top behavioral, financial, and temporal factors that drive conversions:

Factor	Business Insight
Call Duration	Longer calls lead to higher conversions. Each extra second increased the odds of success by 0.42%, showing that meaningful conversations matter.
Previous Campaign Success	Customers who subscribed in a previous campaign are 10x more likely to convert again—retargeting works.
Best Months	March and October had significantly higher conversion odds, while May and July underperformed. Timing is crucial.
Contact Method	If the contact method was "unknown", conversion odds dropped by 80%. Ensuring reliable contact channels is critical.
Loan Status	Customers with housing loans (–44%) or personal loans (–34%) were significantly less likely to subscribe—likely due to constrained liquidity.
Customer Profile	Retirees and students were more likely to subscribe. Those with tertiary education had a 53% higher likelihood to say yes.
Number of Contacts	Over-contacting hurt performance—each extra call reduced success odds by 9.7%. Fewer, well-timed calls are better.

These insights not only validated the model, but also gave us direct levers to act upon in campaign design.

Based on the model, I designed a targeted campaign strategy with multiple levers for improvement:

1. Prioritize Top Leads

Using the model's predicted probabilities, I ranked all customers and focused on the top 10% most likely to subscribe.

Performance Metric	Call All (45K)	Call Top 10% (~4.5K)
Conversion Rate	11.7%	~54%
Net Profit	$15,358	$40,658
ROI	Baseline	2.5× higher
Cost	$90,422	$9,042

Calling just the best 10% saved 90% in cost while increasing profitability by over 2.5x.

2. Limit Call Frequency

Limit follow-ups to 2–3 touches. The model showed diminishing returns—and even harm—after multiple contacts.

3. Focus on Human Connection

I proposed training representatives to focus on:

• Understanding financial goals
• Using customer success stories
• Guiding without pressure

Calls should feel consultative, not salesy.

4. Use Multi-Channel Reinforcement

Instead of relying only on phone calls, I recommended:

• App notifications
• Reminder emails
• SMS nudges

These gentle reinforcements keep the offer top-of-mind without overwhelming the customer.

5. Build a Referral Engine

Referral programs are a low-cost acquisition tool. A simple incentive—like $25 bonus to both parties—encourages word-of-mouth without major upfront investment.

This approach isn't just efficient—it's scalable and repeatable:

• Profit increased 2.5x when using predictive targeting
• Cost per acquisition dropped dramatically (90% cost reduction)
• Customer goodwill increased, due to fewer spam calls and more relevant offers
• As new leads arrive, the same model can assign a score—making campaign planning almost automatic

This strategy strengthens customer trust, marketing efficiency, and bottom-line profitability while creating a repeatable framework for future campaigns.

Key Learnings

• Data-driven targeting dramatically outperforms mass marketing approaches
• Customer context (timing, financial situation) is as important as demographics
• Quality of interaction (call duration, personalization) strongly influences conversion
• Simple interpretable models can provide substantial business value when applied strategically