Evidlake
Menu

Explainable Scoring

Also known as: interpretable scoring, score explanation, reason codes.

Returning, with every score, the reasons behind it: which rules fired, what each criterion contributed and what data were missing.

Updated Last reviewed

Synthetic dataExamples on this page use synthetic data created for illustration.

Not adviceThis is educational material. It is not investment, business or other professional advice.

Definition

Explainable scoring returns, with every score, the reasons behind it: which rules fired, how much each criterion contributed, and which data were missing. A person can then check the result, contest it, and see what would change it.

In plain terms

A score alone invites trust or distrust without evidence. An explanation lets a reader check the arithmetic, notice a wrong input and understand what to change. The explanation must be faithful: it has to describe the calculation that actually produced the score, not a plausible story told afterwards. This is why simple, open models are often preferred where decisions matter; explanations added after the fact to an opaque model may not reflect how it really decided.

An explanation should also carry the version of the rules and the data it used, so that the same score can be reproduced later.

Example

A candidate scored with the weighted model from Weighted Scoring Model (rules version v3, evaluated 2026-10-20):

CriterionWeightRatingContribution
Price0.50.80.40
Reviews0.30.60.18
Delivery0.21.00.20
Score0.78

“What would change it?” is answerable: lifting the reviews rating from 0.6 to 0.9 would add 0.09 and give 0.87. A bare “0.78” supports no such question.

Common misunderstandings

  • “Any story about the score counts as an explanation.” It must match what the calculation really did.
  • “Explainable means simple.” A rule set can be large; what matters is that every step can be traced.
  • “Explanations are only for regulators.” They are how users, reviewers and the builders themselves find mistakes.
  • Weighted Scoring Model

    Rates each candidate on several criteria, multiplies each rating by an agreed weight and adds them up. Trade-offs become explicit, and the answer depends on the weights.

  • Three-valued Logic

    Logic with a third value, unknown, besides true and false. A rule that depends on a missing value is unknown, not false.

  • Reproducibility

    A result is reproducible when someone else, with the same data and code, gets the same result. It needs the data version, code, settings and environment to be recorded.

References

  1. Cynthia Rudin. Stop Explaining Black Box Machine Learning Models for High Stakes Decisions and Use Interpretable Models Instead . Nature Machine Intelligence, 2019. DOI: 10.1038/s42256-019-0048-x
  2. Christoph Molnar. Interpretable Machine Learning . Online book.