2. Predictive Analytics

Decision Metrics and Thresholds

Translate scores into actions using consequences, treatment effects and capacity

Decision Metrics and Thresholds

A score does not act

A threshold creates false positives and false negatives. Under simple constant costs, choose positive when

pCFN>(1p)CFP,pC_{FN}>(1-p)C_{FP},

so

p>CFPCFP+CFN.p>\frac{C_{FP}}{C_{FP}+C_{FN}}.

This formula assumes calibrated pp, mutually exclusive actions and correctly specified costs.

Intervention value needs an effect

Suppose retaining a would-be churner is worth £50 and a message costs £8.

  • If the message certainly prevents churn, act when p(50)>8p(50)>8, or p>0.16p>0.16.
  • If it prevents only 20% of churn, act when p(0.20)(50)>8p(0.20)(50)>8, or p>0.80p>0.80.

Risk is not treatment responsiveness. Targeting the highest-risk customers can waste capacity if they cannot be influenced.

Threshold audit

Py

Compare thresholds by decision cost

Idle

The sample is deliberately small. In practice, choose the threshold on validation data and report uncertainty on a later test period.

Capacity changes the policy

If HarborMart can contact only 1,000 customers, the policy may be “select the top 1,000 eligible expected incremental values,” not “score above 0.5.” Capacity, contact fatigue, fairness and channel constraints enter the ranking.

Compare policies at equal capacity:

PolicyCustomers contactedExpected incremental contribution
random eligible1,000baseline
highest churn risk1,000depends on effect overlap
highest uplift1,000depends on credible causal model
constrained uplift1,000effect minus cost with policy constraints

Decision curves, not metric shopping

Plot or table value across plausible thresholds, costs and capacities. A model that wins at one threshold may lose elsewhere.

AUROC averages ranking across thresholds the business may never use. F1 assigns a particular symmetric trade-off without pounds, capacity or intervention effects. Use them as diagnostics, not objectives by default.

Quick check

A model’s probabilities are perfectly calibrated overall but systematically too low for one customer group. Can the same economic threshold be applied safely?

Answer
Not without subgroup and impact analysis. Overall calibration can hide group error. Reassess data, calibration, consequences, fairness obligations and whether a group-specific repair is lawful and substantively justified.

Next: Explainability and Drift

Copyright © 2026