🤖 machine learning

How Simple Can It Get? From Interpretable Equations to Readable Rules for Financial Decision Making

This paper proposes a framework for progressively simplifying interpretable financial models from complex equations into readable rules and scorecards, demonstrating that such simplifications can significantly improve human readability with minimal loss in predictive performance while providing theoretical bounds to anticipate fidelity degradation.

Original authors: Adia Lumadjeng, Ilker Birbil, Erman Acar

Published 2026-08-11

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Adia Lumadjeng, Ilker Birbil, Erman Acar

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a mystery, but your only clue is a massive, tangled ball of yarn that contains every single piece of evidence you've ever gathered. In the world of artificial intelligence, this "ball of yarn" is a complex computer model that makes predictions about things like whether someone will pay back a loan or if a transaction is fraudulent. While these models are often incredibly good at guessing the right answer, they are usually so complicated that no human can understand how they reached that conclusion. In the financial world, this is a big problem. Laws often require that if a bank says "no" to your loan application, they must be able to explain why in a way a regular person can follow. If the explanation is a jumble of math that looks like alien code, it's useless.

This is where the idea of "interpretable" models comes in. Think of these as models built to be transparent from the start, like a clear glass box instead of a black one. But here's the twist: even a clear glass box can be so full of gears and levers that it's still impossible to read. The paper you are about to read tackles a very specific question: If we already have a model that is technically "understandable" (because it's just one big math equation), how much can we simplify it before it stops making sense? Can we chop off the confusing parts to make a simple "if-then" rule, like a recipe, without ruining the model's ability to predict the future? The researchers wanted to find the sweet spot where a rule is simple enough for a human to read, but still smart enough to make good decisions.

The Great Equation Shrink

Imagine you have a magical recipe for predicting whether a loan applicant will default (fail to pay back). The original recipe, learned by a computer, is a single, massive equation. It multiplies together 42 different ingredients—like "total interest received," "time since last payment," and "annual income"—each raised to a specific, weird power. It's like a smoothie made of 42 fruits, where the computer decided exactly how much of each fruit to blend.

The problem? If you show this 42-ingredient smoothie recipe to a bank manager, they will stare at it in confusion. It's technically "readable" because it's just math, but it's too long to actually use. The authors of this paper asked: "What if we start with this complex recipe and try to make it simpler, step by step, to see how much flavor we lose?"

They didn't just guess; they took the original equation and systematically stripped it down in four different ways, like peeling an onion:

The Pruned Monomial: They took the recipe and threw away the ingredients that didn't matter much. If an ingredient only changed the taste by a tiny bit, they removed it entirely.
The Directional Rule: They kept the ingredients but stopped worrying about how much of each to use. Instead, they just noted the direction. For example, "If interest is high, risk goes up." They threw away the exact numbers and just kept the "up" or "down" arrows.
The Scorecard: They turned the ingredients into simple Yes/No questions. Instead of "Total payment is $450," they asked, "Is the total payment low?" If yes, you get 5 points. If no, you get 0. They assigned integer points (like 1, 2, or 5) based on importance.
The Tally: This was the simplest version. It was just a checklist. "Did you meet condition A? Yes. Condition B? Yes." You just count how many boxes you checked. No points, no weights, just a simple count.

The Surprising Results: What Gets Lost?

The team tested these simplified rules on four real-world financial datasets, including loan defaults and credit card fraud. Here is what they discovered, and it might surprise you.

1. Cutting the fat is almost free.
When they removed the weak ingredients (the ones with tiny effects) to create the "Pruned Monomial," the model barely lost any accuracy. In fact, on one dataset, they cut the number of features from 42 down to just 7, and the model's ability to predict defaults dropped by a tiny, almost invisible amount (from 0.886 to 0.885). It's like realizing your smoothie tastes exactly the same even if you throw out the 40th fruit. You can make the recipe much shorter without losing the flavor.

2. The "Up/Down" rule is a smart shortcut.
When they switched to the "Directional Rule" (keeping the ingredients but only noting if they go up or down), the results were fascinating. On some datasets, the model stayed incredibly accurate. On others, the model's ranking of who is risky changed a bit, but it still did a great job at the actual job of saying "yes" or "no." This suggests that sometimes, knowing the direction of a risk is enough, even if you don't know the exact math behind it.

3. Turning numbers into "Yes/No" is risky.
This is where things got tricky. When they turned the continuous numbers (like exact dollar amounts) into simple "Yes/No" conditions for the Scorecard and Tally, the models often crashed. On the loan dataset, the accuracy plummeted from a strong 0.885 down to a weak 0.329. It was like trying to describe a complex painting using only black and white stickers; you lose too much detail. The paper suggests that for some types of data, the exact number matters more than just knowing if it's "high" or "low."

4. Predicting the crash before it happens.
One of the coolest parts of the paper is that the authors found a way to predict how well the "Directional Rule" would work before they even built it. By looking at the original math equation and how the ingredients relate to each other, they could calculate a score that told them, "Hey, if you throw away the numbers and just keep the directions, your model will still be about 75% faithful to the original." It's like being able to guess how a cake will taste just by looking at the list of ingredients, without even baking it.

What Humans Actually Think

The researchers didn't just look at the math; they asked real people—finance experts and AI researchers—to read these simplified rules and tell them which ones they liked.

The results were a tale of two different worlds.

Finance and Risk Experts loved the Directional Rule. They preferred it 57% of the time. They liked the idea of a clear sentence: "Predict default when interest is high and payments are low." It felt natural to them.
AI and Machine Learning Researchers preferred the Scorecards and Tallies (the point-based systems). They chose these 72% of the time. They liked the structure of adding up points, which reminded them of traditional scoring systems.

Interestingly, almost no one liked the original, full 42-ingredient equation. It was too messy for everyone. The study suggests that while simplifying a model makes it easier for humans to read, different professions prefer different kinds of simplicity.

The Bottom Line

This paper teaches us that making a model "simple" isn't just about making it shorter. It's a trade-off.

You can often cut out the weak parts of a model for free.
You can sometimes drop the exact numbers and just keep the "up/down" direction without losing much accuracy.
But if you turn everything into simple "Yes/No" boxes, you might lose the ability to make good predictions, especially if the data is complex.

Most importantly, the authors found that a simpler rule doesn't always have to be a perfect copy of the original complex model to be useful. A rule can be a bit "unfaithful" to the original math but still be a very effective classifier. It's like a map: you don't need every single tree and rock drawn on it to get to your destination; you just need the main roads.

The study concludes that in the regulated world of finance, we shouldn't just assume that a model is good because it's "interpretable." We have to measure exactly what we lose when we make it readable. Sometimes, the loss is tiny; sometimes, it's huge. And the best way to explain a decision depends on who is listening.

Technical Summary: From Interpretable Equations to Readable Rules for Financial Decision Making

Problem Statement
In regulated domains like finance, models must be explainable to be deployed. However, a tension exists between formal interpretability and practical readability. While "interpretable-by-design" classifiers (such as ECSEL) produce transparent equations, these equations often involve dozens of features and complex exponents, rendering them cumbersome for human decision-makers, auditors, or regulators to inspect. Existing literature often focuses on constructing models that remain transparent while retaining performance (the "forward" direction). This paper addresses the "reverse" direction: given an already-interpretable model (specifically a single-term monomial), how far can its representation be simplified to improve readability while preserving predictive performance and fidelity to the original model?

Methodology
The study begins with a fitted single-term monomial classifier derived from the ECSEL framework. The model takes the form $z(x) = \alpha \prod_{j=1}^m x_j^{\beta_j}$ , passed through a sigmoid. The authors decompose the information within this equation into four components:

Feature Support: Which features participate.
Feature Values: Whether features are continuous or binary.
Direction: Whether a feature increases or decreases risk (sign of the exponent).
Magnitude: The strength of the feature's influence (absolute value of the exponent).

The authors systematically simplify the monomial by removing these components to generate four distinct readable representations:

Pruned Monomial: Retains continuous values and exact magnitudes but removes features with the smallest exponent magnitudes (pruning).
Directional Rule: Retains continuous values and direction but replaces all exponent magnitudes with their signs (effectively setting all magnitudes to 1).
Scorecard: Binarizes feature values (using median cutpoints) and assigns integer point values based on rescaled exponent magnitudes.
Tally: Binarizes feature values and counts satisfied conditions equally (all points = 1).

The study evaluates these representations across four financial datasets (Loan, Default, FraudEcom, Creditcard) using three metrics:

Predictive Performance: Measured by PR-AUC (Precision-Recall Area Under Curve).
Fidelity: Measured by tie-corrected Kendall's $\tau_b$ between the simplified score and the original full monomial score.
Calibration: Measured by Expected Calibration Error (ECE) after isotonic regression.

Additionally, the authors derive a theoretical bound (Proposition 3.1) to predict the rank fidelity of the directional rule using only the fitted exponents and feature covariance, without needing to evaluate the simplified rule. A human assessment involving finance/risk professionals and AI/ML researchers was conducted to gauge perceived readability and practical preference.

Key Contributions

Controlled Simplification: The paper derives readable rules directly from a learned monomial, isolating the cost of removing specific information types (support, values, direction, magnitude).
Quantification of Simplification Costs: Empirical results show that pruning is "nearly free" in terms of predictive loss. However, binarizing feature values (scorecards/tallies) incurs significant predictive costs, whereas removing effect magnitudes (directional rules) is often cheap.
Predictive Fidelity: The authors demonstrate that the fidelity of a directional rule can be predicted before simplification using the fitted model's exponents and feature covariance, allowing practitioners to anticipate the impact of sign-flattening.
Human Assessment: The study reveals that while simplification improves perceived ease of understanding, practical preferences vary by professional background. Finance/risk professionals prefer directional rules, while AI/ML researchers prefer point-based forms (scorecards/tallies).
Robustness Check: The authors compare post-hoc pruning with sparsity imposed during training (Iterative Hard Thresholding), finding that both methods yield nearly identical feature sets, suggesting the extracted structure reflects data learning rather than pruning artifacts.

Results

Pruning: Reducing the feature set (e.g., from 42 to 7 features in the Loan dataset) resulted in negligible PR-AUC loss (e.g., 0.886 to 0.885) and high fidelity ( $\tau_b \approx 0.85$ ).
Binarization: Converting continuous values to binary conditions (Scorecard/Tally) caused substantial drops in PR-AUC (e.g., Loan dropped from 0.885 to ~0.30) and low fidelity, indicating that continuous values carry significant predictive information in these datasets.
Directional Rules: These often maintained high predictive performance despite lower fidelity to the original ranking. For instance, on the Default dataset, the directional rule matched the full monomial's PR-AUC (0.381 vs 0.380) despite a fidelity of 0.744. This highlights that fidelity can erode faster than predictive performance.
Prediction Accuracy: The proposed method for predicting directional rule fidelity was highly accurate for most datasets (Mean Absolute Error $\approx 0.017$ ), though it struggled with low-cardinality count variables (FraudEcom) where ties affected the correlation-to-rank conversion.
Human Perception: All simplified forms were perceived as easier to understand than the full pruned monomial. However, the "best" representation is subjective; finance professionals favored the directional rule (57% preference), while AI/ML researchers favored point-based systems (72% preference).

Significance and Claims
The paper argues that in financial decision-making, an interpretable model is not necessarily a readable one. It posits that readable explanations should be treated as measurable reductions of an interpretable model rather than assumed faithful representations simply because they are easier to read.

The authors claim that:

Simplification is not a single ladder of decreasing accuracy; the cost depends on where the predictive information resides (e.g., in feature support vs. continuous values vs. exponent magnitudes).
Simpler rules can remain effective classifiers even if they do not faithfully reproduce the original model's ranking, suggesting a divergence between predictive utility and model fidelity.
The transition from complex equations to human-readable rules involves trade-offs that can be quantified and, in the case of directional rules, anticipated prior to simplification.
Practical preferences for these rules are not universal but depend heavily on the professional background of the decision-maker.

The study concludes that while simplification improves perceived readability, it must be balanced against the specific information loss required for the task, and that "readable" does not automatically equate to "faithful."

The Great Equation Shrink

The Surprising Results: What Gets Lost?

What Humans Actually Think

The Bottom Line

Technical Summary: From Interpretable Equations to Readable Rules for Financial Decision Making

More like this