How Many Features Should I Have in My Credit Model?

Leland Burns & Jim McGuire
3 days ago
4 min read

“How many features should I have in my model?”

It sounds like a simple question. But like so many modeling decisions—especially in the credit space—the honest answer is: it depends.

We often hear this question from internal teams and executives alike. It comes up when a team is building its first in-house model, when it's looking to upgrade an existing scorecard, or even when trying to explain why their current model looks the way it does.

At Ensemblex, we've built models for consumer lenders across a wide range of products, geographies, and data contexts. We’ve seen feature sets that range from 20 to over 150 variables—each one the right fit for the data and goals at hand.

So how do you know where your model should land?

Here’s how we think about it.

Start with the basics: What’s a model feature?

A feature is a variable the model uses to make a prediction—like the number of recent delinquencies on a credit file, the amount of revolving utilization, or a custom cashflow-derived metric.

Features are derived from raw data. Sometimes, they come pre-assembled in a structured format, like attributes from a credit bureau. Other times, they're custom-built from more complex sources, like tradeline-level credit data or transactional cashflow records.

Regardless of source, features become the model's building blocks. But the number of building blocks you use—and how you use them—depends on a few key factors.

Algorithm choice drives a lot

Different algorithms tolerate different feature volumes.

A traditional logistic regression, for example, performs best when you’re highly selective. Because it’s sensitive to multicollinearity and overfitting, a bloated feature set can erode performance.

But modern machine-learning methods like XGBoost or LightGBM handle this differently. They’re generally more tolerant of feature count, better at managing correlation, and often benefit from larger—but still curated—feature sets.

That means if you’re using an ML method, you can include more features. But just because you can doesn’t always mean you should.

The size of your dataset matters too

No matter what algorithm you’re using, data volume sets an upper bound.

If your dataset includes just a few thousand observations—say, 5,000 funded loans—it’s tough to justify 100+ features. You’ll risk overfitting, noise sensitivity, and unreliable generalization.

In those cases, we often see models perform best with 20–40 features, sometimes even fewer.

On the other hand, if you’re sitting on hundreds of thousands or even millions of observations, you have more flexibility. In those situations, we’ve seen performant models with 100+ features—but only after a careful, methodical curation process.

Feature count depends on the quality and variety of your data

Sometimes we’re given 3,000 raw features from a bureau or partner. But once we remove duplicates, filter on importance, and eliminate high correlations, only a fraction of those make the cut.

Adding data sources helps. If you have both a traditional bureau file and an alternative bureau, plus internal cashflow data, you’ll almost certainly unlock more useful features. Variety increases dimensionality, which gives models more signals to learn from.

Still, there are diminishing returns. Two bureaus that use similar derivation logic might offer overlapping value. A new data source with fundamentally different signals is more likely to expand your usable feature set.

How we reduce features in practice

When we’re not constrained by algorithm or data size—say we’re building an XGBoost model with robust data—our goal is efficiency. We don’t just throw in 200 features and call it a day.

Here’s the simplified path we follow:

Initial filtering with a Boruta-style test We adapt the Boruta algorithm to compare real features to random “shadow” ones. Any feature that performs worse than random noise gets dropped. This quickly trims the fat.
Correlation pruning Highly correlated features don’t add incremental value and increase the complexity of the model. We test across various correlation thresholds, removing the lesser important feature of each pair matched at that threshold. We find the point of diminishing return across these thresholds and use it to further trim the feature set.
Truncating the long tail Many model runs produce a few high-importance features and a long tail of marginal ones. We test how performance holds up if we cut the tail—e.g., comparing the model’s AUC with 50 features vs. 100. If performance is similar, we choose the simpler version.
Manual and compliance review Once we’ve shortlisted the top features, we review them for interpretability, monotonicity, and business alignment. We also loop in compliance and business stakeholders to flag anything problematic or confusing before finalizing the model.

This process typically gets us to a streamlined, high-performing model that’s easy to monitor and explain.

Final thoughts

There’s no one-size-fits-all answer to the question of “how many features.”

But here’s what we know:

The model’s algorithm, data size, and source variety all play a role
Bigger isn’t always better—especially when performance plateaus or complexity grows
A thoughtful curation process makes a big difference in long-term model governance
Even high-performing models benefit from having fewer, cleaner, well-understood features

Whether you’re just getting started or refining a production model, Ensemblex helps internal teams make these decisions with confidence—and with the processes and documentation to back them up.

If you’re wondering whether your current model is the right size for your data, or if you want help streamlining a feature set without losing performance, we’d love to talk.