The Hong Kong University of Science and Technology · FACULTY OF ARTS & HUMANITIES

MAIE5103 Chap.5 Equal Opportunity in Lending Decisions

- one subject, every graph, every model, every mark
8 Chapters3-page Bible
Our own words - no uploaded lecturer files
Updated for this semester
Chapter 5 of 14 · MAIE5103

Equal Opportunity in Lending Decisions

A score is not a decision

The lending lesson is built on Hardt, Price and Srebro's 2016 paper on equality of opportunity in supervised learning. A credit model estimates repayment ability from income, debt and past payments; the bank then sets a cutoff, the score an applicant needs for approval. That choice produces two kinds of mistake with two kinds of harm.

A wrong approval, a false positive, costs the bank a repayment loss.

A wrong rejection, a false negative, costs a good borrower a chance they should have had.

Three rules that need not agree

The lesson opens by asking whether three intuitive rules would always make the same choices: the same score requirement for everyone, the same overall approval rate for each group, and the same approval rate for people who would repay. They need not.

Equal totals can hide unequal chances, because a group's approval rate says nothing about whether its good borrowers or its bad risks are the ones being approved.

Equal opportunity matches the approval rate among people who would repay; equalized odds also matches it among people who would not.

Changing the rule instead of the model

The paper's method is post-processing: keep the score, use past outcomes to find decision rules that meet the chosen check, often with different cutoffs per group, and pick the one with the lowest error cost.

Where fixed cutoffs cannot reach the target, a planned random choice between two cutoffs matches the rates over many decisions. On the paper's credit data, equal opportunity kept 92.8% of the maximum possible profit and the equal-share rule 69.8%, in one dataset under one cost setting. The lesson ends on what no rate check can prove.

In this chapter

What this chapter covers

  • 01

    Credit scores, cutoffs and loan decisions

  • 02

    Wrong approvals and wrong rejections

  • 03

    Why ignoring the group label and equal shares fall short

  • 04

    Equal opportunity for people who would repay

  • 05

    Equalized odds for both kinds of borrower

  • 06

    Post-processing with group-specific cutoffs

  • 07

    Planned randomisation between two cutoffs

  • 08

    The credit-score case and the limits of rate checks

Worked example · free

Read an equal-share claim from a credit union

Q [6 marks]. AskSia-authored practice. A credit union reports that it approved 40% of loan applications from each of two neighbourhoods. Records show that in Neighbourhood E, 30 of 50 applicants who later proved good borrowers were approved, along with 10 of 50 who would have defaulted. In Neighbourhood F, 22 of 50 good borrowers and 18 of 50 who would have defaulted were approved. Check the claim against equal opportunity and equalized odds. The marks shown are an AskSia study allocation, not a University marking scheme.
Each bar is the share of maximum profit a rule keeps in the paper’s credit case.Each bar is the share of maximum profit a rule keeps in the paper’s credit case. Equal opportunity costs about seven points and the same-total-share rule about thirty, on a single dataset and a single set of cost assumptions.No fairness rule100.0%Same cutoff99.3%Equal opportunity92.8%Equalized odds80.2%Same total share69.8%
Each bar is the share of maximum profit a rule keeps in the paper’s credit case. Equal opportunity costs about seven points and the same-total-share rule about thirty, on a single dataset and a single set of cost assumptions.
  • 1Confirm the equal-share figure from the counts.
  • 2Compare approval rates among good borrowers.
  • 2Compare approval rates among those who would default.
  • 1State what the equal-share claim concealed.
Each neighbourhood had 40 approvals out of 100, so the 40% figure is correct and the same-total-share rule is met. Among good borrowers, approval was 30/50 = 60% in E and 22/50 = 44% in F, so equal opportunity fails by sixteen points. Among those who would default, approval was 10/50 = 20% in E and 18/50 = 36% in F, so equalized odds fails on both conditions. The equal share concealed that the credit union's decisions sort Neighbourhood F far less accurately, so its good borrowers are turned away more often while more of its risky applications are approved.
Sia tip — Whenever a lender quotes one approval rate per group, split it into the rate for people who would repay and the rate for people who would not before you accept it.
Glossary

Key terms

Cutoff
The score an applicant needs for approval, chosen by the decision-maker rather than produced by the model.
Wrong Rejection
Turning down an applicant who would have repaid, a false negative that costs the applicant an opportunity.
Wrong Approval
Approving an applicant who will not repay, a false positive that costs the lender a repayment loss.
Post-processing
A mitigation that keeps a trained score unchanged and adjusts the decision rule, often with group-specific cutoffs.
Randomised Cutoff
A rule that chooses between two cutoffs with probabilities set in advance, so that target rates are matched across many decisions.
FAQ

Equal Opportunity in Lending Decisions FAQ

What does equal opportunity mean in machine learning?

It is the fairness criterion proposed by Hardt, Price and Srebro in 2016: among people who truly deserve the favourable outcome, such as borrowers who would repay, each group should receive it at the same rate. It equalises true-positive rates. It does not say what that shared rate should be, and it places no limit on wrong approvals, which equalized odds would add.

Why is approving the same share of each group not enough?

Because equal totals say nothing about who is approved. A lender can approve half of each group while approving most of one group's good borrowers and few of the other's, simply because its score sorts the second group less well. The lending lesson shows exactly this: two groups with identical approval totals and very different chances for the people who would repay.

Why would a fair lending rule use randomisation?

With only one fixed cutoff per group, an equalized-odds target is sometimes out of reach. The paper's method then uses a planned random choice between a lower and a higher cutoff, with the probability set in advance so that approval rates match over many decisions. It does not guarantee any individual's outcome, and the lecture notes it is statistically clean but politically hard to defend.

Study strategy

Assessment move

Practise turning any single approval rate into two conditional rates, one for people who would repay and one for people who would not, using small invented counts. Then say which of the three rules each pair of numbers meets. Finish by writing two sentences on what the paper's credit-score chart does and does not show, because overstating that chart is the easiest way to lose credibility in a written piece.

Working through Equal Opportunity in Lending Decisions in MAIE5103? Sia is AskSia’s AI Arts and Humanities tutor — ask any MAIE5103 Equal Opportunity in Lending Decisions question and get a clear, step-by-step explanation grounded in how MAIE5103 is taught and assessed. Read this chapter free, then take your hardest questions to Sia.

A+Everything unlocked
Unlocks this Bible + all 13 of your subjects at The Hong Kong University of Science and Technology, plus 1,000+ more Bibles across the AskSia library.
Sia - your MAIE5103 tutor, unlimited, worked the way the exam marks it
The full 3-page Bible + practice bank with worked solutions
Chrome extension - sync your LMS so Sia knows your deadlines
Bilingual EN / Chinese on every Bible and every Sia answer
$0.99 Trial
30-day money-back · cancel in one tap · how it works