How I Used Machine Learning to Predict Banking Lead Conversions
I used to work at a fintech company where I got to work closely with banks and financial institutions. And honestly, I was always fascinated by how much effort (and money) they put into acquiring new customers, through ads, referrals, promotions, and more.
Lately, I’ve been wanting to get back into the data game by building something meaningful. And most importantly, I wanted to work on something that actually connects with my background. So I decided to dive into the banking domain once again and explore a simple but important question:
Which channels and customer traits truly lead to conversions, and which ones just drain the marketing budget?
The Idea
The goal was simple:
To help banks identify which customers are most likely to convert and which ones bring the highest lifetime value. This way, they can focus their marketing and sales efforts where it really matters.
Since I didn’t have access to real customer data (because… well, banks don’t share that easily😅), I created a simulated dataset of 20,050 rows and 26 columns, designed to very closely resemble what an actual bank might have.
It included customer demographics, lead sources, engagement metrics, product interests, income, and even a few behavioral indicators like website visits and time spent online.
Tools & Technologies
I used a mix of programming, visualization, and cloud tools to bring this project to life:
Python (with Pandas, NumPy, Matplotlib, Seaborn, Scikit-learn) for data analysis and modeling, Tableau for interactive dashboards, and Azure ML to simulate a cloud deployment.
The Process (Step-by-Step)
I wanted this project to feel practical. I mean something that could genuinely be used in a real bank setting.
So here’s how I approached it.
1. Data Cleaning & Preprocessing
I started by cleaning the dataset, fixing missing values, encoding categorical columns, and scaling the numerical features.
I feel like this step is always underrated, but it sets the foundation for everything that follows. I mean, Garbage in, Garbage out, right?
2. Exploratory Data Analysis (EDA)
Next, I explored the data to understand trends:
- Which channels brought the most leads?
- What kind of customers tend to convert more?
- How does income or province affect conversions?
I visualized relationships between conversion rates and demographics, and started to notice interesting patterns. For example, referrals and walk-ins performed much better than digital ads or partnerships.
3. Channel Effectiveness Analysis
Here’s where it got exciting.
I analyzed how different acquisition channels performed. And that was not just in terms of how many leads they brought, but how well those leads converted.
It turned out that online ads brought volume, but referrals brought value.
I also looked at how location and income affected conversions. Alberta, BC, and Ontario drove the best conversions, while Manitoba and Nova Scotia lagged behind. Most leads earned between $58K–$92K, with conversion rates peaking for those in the $80K–$200K range. Focusing on these high-performing provinces and middle-to-high income segments could make marketing far more efficient. These insights alone could save banks thousands in wasted ad spend!
4. Predictive Modeling
Then came the machine learning part.
I trained a Random Forest model to predict how likely each lead was to convert into an actual customer.
The model picked up some interesting patterns:
- Income and engagement metrics (like website visits and email opens) had the biggest impact.
- Gender and age didn’t matter as much as I expected.
- And the referral source turned out to be one of the strongest predictors overall.
Once the model was trained, I tested it on a new batch of leads to see how well it could generalize. It performed well, accurately predicting which leads would convert into customers.
And it wasn’t just accurate. It was interpretable.
We could actually explain why certain leads were more valuable, which is crucial when sharing insights with marketing and product teams.
5. CLTV (Customer Lifetime Value) Analysis
Next, I built a Linear Regression model to estimate how much long-term value each customer might bring once they convert, which is their Customer Lifetime Value (CLTV).
This helped connect marketing actions to financial outcomes and determine not just who converts, but who stays valuable.
After training, I fed new, unseen customer data into the model to predict each customer’s estimated CLTV. This showed how the model could take raw inputs and generate actionable long-term value estimates for future leads. I also determined the key drivers of lead conversion (using the Feature Importance method.)
6. Azure AI Integration (Simulated)
I also designed the workflow so that it could be deployed on Azure Machine Learning (similar to how banks deploy predictive systems in production.)
Due to subscription limits, I didn’t deploy it fully, but the setup is ready for cloud integration.
7. Tableau Visualization
Finally, I took all the insights to Tableau and built interactive dashboards.
Users can filter by province, age, income, or lead source and instantly see how conversion likelihood and CLTV change across segments.
It’s the kind of visualization that a marketing or strategy team could use right away to make decisions.
🔗 Interactive Dashboard: View it on Tableau Public
💻 Full Project + Code: Check it out on GitHub
What I Learned
This project taught me that data science isn’t just about accuracy, but it’s about clarity. It’s about helping decision-makers see the story behind the numbers.
Some key takeaways:
- Predictive analytics can turn guesswork into an actionable strategy
- Not every high-volume channel drives long-term value
- Sometimes, the “boring” features (like engagement) matter more than flashy ones
And most importantly, that even simulated data can uncover real business lessons when approached thoughtfully.
Wrapping Up
Again, I didn’t include the entire code here (it would get too technical for most readers), but if you’d like to explore how everything works (from data cleaning to model training, and also the dataset), you can check out the full notebook on GitHub.
This was a fun project that combined my fintech background with my passion for data analytics. And it reminded me that with the right data, even a simulated one, you can tell powerful stories that connect numbers to real-world strategy.
Btw, if you’re working in banking or marketing analytics, I’d love to hear how you’re using data to predict customer behavior! Please feel free to connect with me on LinkedIn or Instagram.
