← Back to projects FINANCE · MACHINE LEARNING · POWER BI

Predicting Credit Default

A k-NN classifier that catches 6 out of 10 defaulters before the money leaves the bank — paired with a Power BI dashboard that shows loan officers exactly where the risk is concentrated.

30,000+ applicants 21.5% baseline default rate 61% recall 89% accuracy
The Challenge

Lending money is inherently risky — lending blindly is what destroys margins. The portfolio I analyzed had a default rate of 21.5%: for every five loans approved, more than one resulted in a loss. The goal was to build both a predictive gatekeeper and a visual risk map loan officers could actually use day to day.

Live Dashboard

Explore the Power BI dashboard

Having trouble viewing it? Open the dashboard directly.

The Process

From raw applications to a risk-ready model

Data profiling & business framing

Loaded 30,000+ loan applicant records — age, income, home ownership, loan intent, loan grade. Before touching any algorithm, I isolated high-risk renters with a history of default and found that medical expenses and debt consolidation were the biggest drivers of losses in that segment.

Cleaning & exploratory analysis

Filled missing interest-rate values (about 10% of rows) with the median, stripped out impossible outliers like ages over 100 and employment lengths over 60 years, then ran a correlation matrix to scout early relationships with default status.

Feature engineering

One-hot encoded the categorical fields, split the data 70/30 for training and testing, and scaled the numeric features — without that step, an income in the hundreds of thousands would have completely drowned out a variable like interest rate.

Choosing the right metric — and the right model

Accuracy is a trap here: a model that approves everyone still looks roughly 80% accurate while the bank quietly bleeds money on the 20% who default. Recall on defaulters was the metric that actually mattered. A scatterplot of income against interest rate showed no clean line between good and bad loans, which ruled out a simple linear approach — so I tested a linear SVM against a distance-based k-NN model.

Translating the model into decisions

A model living in a notebook doesn't change anyone's behavior. I moved the results into Power BI so loan officers could see risk visually — by grade, by home-ownership status, by loan purpose — instead of trusting a single score.

Key Findings

What the dashboard exposed

  • k-NN beat the linear model. The distance-based k-NN classifier caught 61% of actual defaulters at 89% overall accuracy, comfortably outperforming the linear SVM.
  • "G-grade" loans are a black hole. This grade carried an almost 98% default rate — not a calculated risk, essentially a guaranteed loss that should be restricted outright.
  • Renters are the highest-risk segment. A 31% default rate, concentrated in debt-consolidation and medical loans.
  • Homeowners are the safest bet. Just a 7% default rate, making them prime candidates for premium credit products.
Impact

Combining the k-NN model's ability to catch 6 in 10 defaulters with a dashboard that makes risk visible by grade and profile gave the bank a path from reactive debt collection to proactive lending policy — the fix wasn't to stop lending, it was to stop lending blindly.

Next project: FP&A Dashboard for Condominium Management →