← ds learning track
notes · 21

Explainability, Feature Effects & Model Interpretability

A model can perform well and still remain hard to trust if nobody can explain what drives its decisions. Explainability methods try to connect predictions back to feature effects, local decision logic, and business intuition.

Start with global versus local interpretation, then move into feature effect curves, monotonicity, and SHAP-style contribution logic. The goal is not just to open the black box, but to understand what kind of box you are dealing with in the first place.

Performance answers “how good”; explainability answers “why”

Predictive view

Metrics like AUC, KS, Brier, or log loss tell you whether the model is useful. They tell you about ranking, calibration, and error.

But they do not tell you what features are driving the predictions or whether the decision logic aligns with domain intuition.

Performance tells you whether the model works.

Interpretability view

Explainability asks which variables matter, how they matter, whether their effect is monotonic or unstable, and why one specific case received a certain prediction.

That is especially important in regulated environments where decisions need to be communicated and challenged.

Explainability tells you how the model thinks.
Credit-risk relevance: a model may be statistically strong but still unacceptable if its effects are economically implausible, non-monotonic without justification, or impossible to defend to governance stakeholders.

A useful order for learning explainability

01

Start with model type

Interpretability depends on the model family. Scorecards, logistic regression, trees, and ensembles are not explainable in the same way.

02

Then separate global from local logic

Global interpretation asks what usually matters. Local interpretation asks why this particular borrower got this particular prediction.

03

Then inspect feature effect shape

Direction, monotonicity, saturation, and interactions often matter more than simple feature ranking.

04

Then challenge the explanation itself

Not every explanation is stable, faithful, or causally meaningful. Some are only approximations of the original model.

The same model can be explained at two different levels

Global importance tells you what usually matters across the portfolio. Local importance tells you what mattered for one selected case. Those two views can differ sharply.

How a feature changes predicted risk

Partial dependence style views help show the average relationship between a feature and the model output. This is where shape matters: linearity, threshold effects, flattening, or unexpected reversals.

SHAP-style thinking: from base score to final prediction

A local explanation often starts from a base rate and then adds or subtracts feature contributions until the final prediction is reached. This section gives the intuition without pretending the additive decomposition is the same thing as causal truth.

Contribution waterfall

Case

Base PD
Final PD
Largest upward push
Largest downward push
Key idea: local additive explanations are useful for narrative and diagnosis, but they are still a representation of the model logic, not a proof of real-world causality.

Should feature effects be monotonic?

In many regulated settings, monotonic effects are easier to defend. But forcing monotonicity may reduce fit. This section shows the tradeoff between stability, interpretability, and local predictive gain.

Explainability tools compared

Tool / view Main question Strength Main caution
Coefficients / scorecard pointsHow does the model move risk globally?Direct and transparentMostly linear logic
Feature importanceWhich variables matter most overall?Fast global summaryCan hide effect direction and interactions
Partial dependenceHow does average prediction move with one feature?Good effect-shape intuitionCan mislead under strong correlations
Local contributions / SHAP-styleWhy this specific prediction?Good case-level narrativeStill an approximation / representation
Monotonic constraintsCan the model be made more defendable?Improves governance and stabilityMay reduce fit or oversimplify
Surrogate explanationCan a simpler model mimic the complex one?Useful for communicationExplains the surrogate, not necessarily the original exactly

Concepts every validator should keep

global vs local

These are different questions

Global importance asks what matters on average. Local explanation asks what mattered for one specific case.

effect shape

Direction is not enough

Two models can rank a feature similarly while implying very different shapes: linear, saturating, threshold-like, or non-monotonic.

correlation

Correlated features complicate interpretation

When variables overlap heavily, contribution allocation can become unstable and importance rankings can be misleading.

faithfulness

An explanation can be intuitive but incomplete

Some explanation tools are excellent for communication but do not perfectly represent the original model’s internal mechanics.

governance

Explainability is partly a governance problem

The right level of explanation depends on model type, regulatory expectations, decision materiality, and stakeholder needs.

causality

Explanation is not causation

A feature can be influential in the model without being the true causal driver in the underlying economy or borrower behavior.

What to leave this page with

Explainability is not one thing. It includes global understanding, local reasoning, effect shape analysis, and governance-friendly constraints.

The useful order is: first identify the model family, then separate global from local interpretation, then inspect feature effect shapes, then evaluate whether the explanation is stable, faithful, and defendable enough for the use case.

Once that structure is clear, model interpretation stops being a cosmetic add-on and becomes part of how the model is actually trusted.