Decision Model Evaluation — Patterns

Choose the operating point from a stated cost ratio

Problem. A threshold gets chosen because it maximised some aggregate figure on a validation split, and the figure weighted both kinds of error equally. In the product they are not equal — a wrongly escalated request costs minutes of a person's time and a wrongly auto-resolved one costs a customer.

Approach. Write down the relative cost of each error type before looking at any curve, in whatever unit the business already uses. Sweep the threshold, compute expected cost per thousand requests at each point using the production base rate, and pick the minimum. Record the ratio and the base rate alongside the chosen threshold, so that when either changes it is obvious the threshold must be revisited.

Trade-off. Eliciting a cost ratio forces a conversation nobody wants to have, and the number agreed on is usually a guess dressed as a figure. It is still better than the implicit ratio of one-to-one that an unexamined threshold encodes, but it invites false precision in a decision that is genuinely a judgement.

Report at the production base rate, always

Problem. Evaluation sets get rebalanced so the rare class is well represented, which is reasonable for training and for comparing models. Precision computed on that set is then quoted as though it described production, where the rare class is a fiftieth as common, and the real precision is far lower.

Approach. Keep the rebalanced set for discrimination comparisons, and maintain a second evaluation population that preserves the production prevalence of every class, against which all reported precision, alert volume and workload figures are computed. Where a stratified sample is necessary to get enough rare-class examples, carry the sampling weights through and reweight before reporting. State the prevalence on the face of every report.

Trade-off. A prevalence-preserving set needs to be large to contain enough rare-class examples for a stable figure, which makes it expensive to label and slow to refresh. Two populations also mean two sets of numbers and a standing risk of the wrong one being quoted.

Separate the ranking from the cut

Problem. Two models are compared on a single accuracy figure at a fixed threshold. The one that loses may have better ranking throughout and merely need a different cut; the one that wins may be worse everywhere except at that one point.

Approach. Evaluate in two stages. First compare how well each model orders cases, using a threshold-independent summary, to decide which model has more signal. Then, for the chosen model, select the operating point separately against the cost ratio and base rate. Keep the two decisions in separate reports, because they answer different questions and have different failure modes.

Trade-off. A good ranking does not guarantee a usable operating point: a model may order cases beautifully and still have no threshold at which both error rates are acceptable. Threshold-independent summaries are also misleading on heavily skewed data, where the figure is dominated by the abundant class and a precision-focused summary is the better instrument.

Treat abstention as a scored outcome

Problem. Forcing a decision on every case means the model must guess on the fraction where it has no signal, and those guesses are concentrated exactly where the cost of being wrong is highest. Scoring the model as if it had to answer everything hides how much of its accuracy comes from the easy majority.

Approach. Allow a third outcome — defer — and report accuracy on the decided subset alongside the deferral rate and the composition of what was deferred. The pair of numbers makes the real trade visible: a model deciding seventy percent of cases at very high accuracy may be worth far more than one deciding everything at middling accuracy, once the cost of the human queue is counted.

Trade-off. Deferral moves work rather than removing it, and a model that learns to defer whenever it is unsure can push an unbounded queue onto a team with fixed capacity. The deferral rate therefore needs a ceiling of its own, and that ceiling is a staffing decision rather than a modelling one.