Selling Smarter: Predicting Customer Conversion in Bank Telemarketing Campaigns with Machine Learning

May 22, 2023

📊 The Business Problem

Banks regularly run telemarketing campaigns to promote products like term deposits. But success rates are typically low—most people hang up, and agents waste time calling unlikely prospects. I asked: Can we use data to predict which customers are likely to subscribe, and prioritise those calls?

This project tackles exactly that using a real-world dataset of Portuguese bank telemarketing calls.

💡 Project Goal

Telemarketing remains a common channel in the banking sector, but with low conversion rates and high resource costs. My objective was clear: Build a machine learning model to predict whether a customer will subscribe to a term deposit based on call and demographic data, and make the process smarter, faster, and more efficient.

🔍 Step 1: Exploratory Data Analysis (EDA)

Using a real dataset of Portuguese bank marketing campaigns, I explored the relationships between customer attributes and subscription behavior.
Key findings:

Call duration is by far the strongest indicator of subscription—longer conversations increase likelihood.
Month of contact matters: May has high call volume but lower success, whereas October shows better conversion rates.
Job type and education level subtly influence outcomes, hinting at financial awareness and risk appetite.

I visualised distributions, correlations, and campaign effectiveness to guide feature selection and model strategy.

🧼 Step 2: Preprocessing & Feature Engineering

To prepare the dataset for machine learning:

Cleaned nulls and removed leakage features (e.g., duration was used with caution).
Categorical variables were encoded using One-Hot and Label Encoding.
The target variable (subscription) was binary encoded.
I addressed class imbalance using SMOTE and stratified sampling to avoid bias in model predictions.
Features were scaled to ensure consistent weight across models like Logistic Regression.
No use was made of previous campaign outcomes (poutcome), focusing instead on real-time customer features.

Step 3: Model Building & Tuning

I trained and compared multiple models:

Logistic Regression for interpretability.
Random Forest for robustness and feature importance.
Neural Network for high performance on tabular data.

After testing across cross-validation folds, 👉 Random Forest Classifier emerged as the best performer with:

Accuracy: 89%
Recall (True Positive Rate): 75%
Precision: 87%
F1-Score: 80%

These results strike a good balance, correctly identifying subscribers without over-predicting and wasting agent effort.

📈 Insights for Strategy

The model does more than predict—it reveals strategy:

When to call: Focus campaigns on specific months and customer job segments with higher success rates. May and October are peak months.
Monitor and optimise call durations—but don’t rely solely on them as they’re influenced post-contact.
Who to target: Retired clients and students show higher conversion rates.
What to avoid: Short calls and clients with negative outcomes from past campaigns.
Data shows potential for targeted pre-qualification before dialling begins.

These insights help craft more effective campaigns, not just more accurate ones.

🔧 What I Learned

This project sharpened my skills in:

End-to-end pipeline creation (EDA ➝ preprocessing ➝ modelling ➝ evaluation).
Feature encoding and handling imbalanced classes.
Model tuning and selection based on practical metrics like recall and precision, not just accuracy.

Next Steps

Deploying this model into a Streamlit dashboard for use by marketing teams.
Integrating real-time prediction based on live call data.
Testing model generalisation on newer customer data.

Final Thoughts

This project proves machine learning can turn noisy marketing campaigns into precision targeting tools. By building a pipeline that spans from raw data to business insight, I’ve created a solution that could help real organisations boost ROI and improve customer experience.

💬 Want to collaborate, or see the code and visuals in action? [Let’s talk.]

Check it out!

Portfolio.