We Underwrite with a Lot of Rules. Can We Safely Get Rid of Them?

One of the most common questions we hear about credit model modernization is this:

“We use a lot of underwriting rules—how can we safely reduce or replace them?”

Or more bluntly: “Our model isn’t doing much. Most of the real work is happening in the knockouts.”

Many lenders start with a simple rules-based credit policy, planning to grow out of it over time. But shedding the knockouts is easier said than done. What starts as a short-term solution often turns into a long-term limitation, adding complexity and fragility.

At Ensemblex, we’ve helped lenders of all sizes transition from rules-heavy systems to modern, model-driven credit strategies. In this post, we’ll explain:

• Why you likely have too many rules

• Why even “working” rules can be a problem

• How to roll them back safely

First, What Do We Mean by “Rules”?

In this post, we’re talking about knockout rules—hard-coded decisions based on a single variable or threshold. Examples include:

• No bankruptcies in the last 5 years

• No 60+ day delinquencies in the past 12 months

• Minimum income of $X

These rules often make sense on their own. But over time, they tend to accumulate—sometimes in the dozens—creating complexity and rigidity in your credit process.

Why Do Lenders Rely on Rules?

1. Startup simplicity

Rules are quick to implement and easy to understand. Many lenders start with a rules-based system while building data volume for model development.

2. Compliance or operational constraints

Some rules reflect hard eligibility constraints (e.g. age, geography, employer affiliation).

3. Unstable inputs

Certain variables—like marketing channel or device type—may be too volatile for model inclusion. But lenders still want to use them to shape approvals.

4. Bad model compensation

Many rules are layered to “fix” problems with an underperforming model. When the model misses risk, operators patch with new rules instead of retraining.

5. Legacy layering

Removing rules must be done carefully, so without regular pruning, they pile up.  Some persist long after their original rationale is forgotten.

What’s the Problem with Rules?

First, we should emphasize that not all rules are bad and many are necessary. Reasons 2 and 3 from our list above, for example, should be kept in place. But for other rules, even if they “work,” they come at a cost.

1. Operational drag

Tuning your credit policy becomes a nightmare. Want to widen approvals? You now need to edit dozens of rule thresholds and interactions instead of adjusting a single model cutoff.

2. Weaker credit decisions

Rules ignore interactions. A statistical model can weigh overlapping signals, prioritize what matters, and rank risk continuously. Rules operate in isolation, leading to less powerful decisions.

3. Biased model training

Here’s the big one: when you knock out swaths of applicants before modeling, your model becomes blind to those segments. That means:

• Your model can’t learn how risky they actually are

• You lose the ability to later expand safely into those segments

• You fall into a dangerous cycle: bad model → more rules → even worse model

This is how rules become a trap.

Our Playbook: How to Roll Back Rules Safely

We never recommend eliminating rules all at once. The key is a measured, iterative approach built on good analytics and smart testing.

Step 1: Map and analyze your rules

Start by building a comprehensive waterfall of your current rules. But don’t just list them sequentially—analyze:

• How many applicants are affected by each rule

• How many are already caught by earlier rules

• How many would be caught by your model alone

This often reveals redundancies. Some rules do no incremental work—they’re overlapping with others or the model itself. These are prime candidates for removal.

Step 2: Prioritize the rest for rollback

Next, sort the remaining rules into priority tiers based on:

• Model coverage: If your model already uses rich features related to a rule, it’s likely safe to remove. For example, if your model includes dozens of delinquency metrics, it probably already captures most of what a 60+ DPD rule does.

• Business sensitivity: Some rules block high-value segments. Start where the business payoff is greatest.

• Historical variability: If the rule has been inconsistently applied in the past, it creates a natural test group, which makes analysis easier.

Step 3: Use smart testing to reduce risk

Once you have a prioritized list, you can begin with a deliberate test strategy to attack it.

Often, we’ll run shaped tests—removing the rule only for applicants with very high model scores. This limits exposure while generating new data. For example, if you’re removing a bankruptcy rule, you might start by approving only previously bankrupt applicants with top-decile model scores. Over time, expand as performance data supports it.

If third-party data is available, you can also supplement with retro bureau studies—pulling risk outcomes from other lenders on similar populations to further de-risk the decision.

Step 4: Monitor performance with visibility into rule lift

When you roll back a rule, track outcomes not just overall, but within the previously excluded segment.

In model monitoring, segment your KPIs (e.g., first payment default, loss rate) by:

• Model score tier

• Rule history (prior KO vs general population)

This shows whether the model is successfully differentiating risk within the formerly excluded population, and helps you decide whether to expand further.

Final Thoughts

Rules aren’t bad. Some are essential. But many are unnecessary and counterproductive.

Transitioning from rules-heavy decisioning to a model-driven strategy brings:

• Simpler operations

• Improved underwriting outcomes

• Better models with richer training data

• More flexibility for future policy changes

At Ensemblex, we’ve helped lenders shed legacy rules without exposing themselves to blind spots or runaway losses. If your current underwriting depends heavily on rules—or you’re not sure whether your model is doing the real work—we’d love to help.

Let’s build something better.