Exam-style
A lender's approval model scores 96 per cent on the data it was trained on and 94 per cent on held-out validation data. Sliced by customer cohort, one group sits at 71 per cent. That cohort is 2 per cent of the training set. What has gone wrong, and what fixes it?
Reveal the answer
D. The cohort is under-represented in the training data. Collect more examples of it, or reweight the ones there are, and re-evaluate per cohort.
Read the two headline numbers first. Training is 96, validation is 94, and a two-point gap that small says the model generalises about as well as it fits. Overfitting produces the opposite shape, near-perfect on the rows it saw and much worse on rows it did not, so reducing capacity here removes ability the model is already using well. Underfitting produces high error in both figures at once, and 96 on training is not that, so more epochs buy nothing. Both of the fit repairs are answers to a problem the numbers rule out. The failure is visible only when the same comparison is run per cohort. A group that contributes 2 per cent of the training rows contributed almost nothing to the parameters, so the model learned the majority’s pattern and applies it to everybody. Its 71 per cent is a data problem showing up as inaccuracy for one set of people, and the repairs that work are on the data side: gather more examples of that cohort, reweight the ones already there so they carry proportionate influence, and check the result cohort by cohort rather than on the average. Dropping the records is the worst move available, because it removes the evidence of the harm while leaving the harm in production. Calling it noise fails on the same ground: 23 points below the overall figure is a pattern, and it is worth confirming on a fresh sample rather than dismissing. The finding belongs in the model’s documentation, next to the headline accuracy, so that whoever approves the release sees both.
Q. 96 per cent on training, 94 per cent on validation, 71 per cent for a cohort that is 2 per cent of the training data. Overfitting, underfitting, or something else?
A. Something else. The cohort is under-represented, so collect more of its examples or reweight the ones you have, then re-measure cohort by cohort.
Why? The diagnostic is one comparison read twice. The gap between training and validation error tells you variance: a wide gap is overfitting, and the repair is less capacity or regularisation. The level of both together tells you bias: high on training as well as validation is underfitting, and the repair is more capacity, better features, or longer training. Here the gap is two points and the level is high, so the model is fitted about right and neither repair applies. Running the same comparison per cohort is what exposes the third cause. A group holding 2 per cent of the rows barely moved the parameters, so the model never had enough examples to learn its pattern, and 71 per cent is what that looks like from the outside. This is how the material frames the effects of bias and variance: not as a training curve, but as effects on demographic groups, where inaccuracy concentrated on one group of people is a fairness finding rather than a tuning defect. Record it as one, in the model card alongside the headline accuracy, and keep the per-cohort slice in the monitoring after release, because representation drifts as the customer base changes.