Risk models, calibration, and delayed outcomes

Choose a target that matches the action

A model target defines what the system learns to predict. Confirmed unauthorized use, a dispute, a default, and a suspicious-activity disposition are different labels. Choose the target from the intended decision and available evidence.

Specify the prediction horizon, population, exclusions, and label maturity. A payment-fraud model should not silently use merchant service disputes as equivalent fraud labels. A credit model should distinguish a technical payment failure from inability to repay. Write a target contract before selecting the algorithm. Otherwise, a sophisticated model can optimize the wrong outcome very efficiently.

Inside the mechanism. Write the target as an observable outcome with a horizon, unit, amount treatment, exclusions, and maturity rule. A dispute occurrence, confirmed unauthorized payment, and final loss are different targets. Match the intervention to the mechanism the target represents. If the action cannot affect the outcome, a predictive model may still be useful for planning but does not establish preventable loss.

A concrete example. A model trained to predict disputes does not automatically estimate unauthorized fraud or preventable loss. The label should match the outcome and action being evaluated. The rule flags 716 of 24,600 mature transactions. Of those flags, 354 meet the synthetic target, giving 49.44% precision. It misses 89 target events. Under the stated cost assumptions, residual loss and operating friction total $17,579. The important result is the connection between the population, action, capacity, and outcome—not one isolated score.

When the assumption fails. A business reports all disputed value as fraud prevented by its model. Define the target, maturity window, exclusions, and intervention assumptions separately. The following worked sequence shows the reference condition, a stress condition, and a response condition with explicit synthetic data. These are comparative assumptions, not measured causal effects.

Follow a worked case3 conditions · 36 figures

A model trained to predict disputes does not automatically estimate unauthorized fraud or preventable loss. The label should match the outcome and action being evaluated.

Target and intervention: reference condition figure preview

01Reference conditionTarget and intervention

Target and intervention: failure and stress figure preview

02Failure and stressTarget and intervention

Target and intervention: controlled response figure preview

03Controlled responseTarget and intervention

Choose a target that matches the action Choose a target that matches the action — the distinction These concepts answer different questions. Read each definition in the context of the section. Proxy label Available outcome related to the concern Intended target Specific event the decision should predict — These concepts answer different questions. Read each definition in the context of the section. Chapter sources · Open image

Choose a target that matches the action Target contract Fictional teaching record. Different cause. Target contract Illustrative data; not a real customer record or a prescribed policy. Outcome confirmed unauthorized payment Defined label Horizon specified maturity window Comparable observation Excluded technical duplicates Different cause Different bad outcomes need different controls — Fictional teaching record. Different cause. Chapter sources · Open image

Choose a target that matches the action Choose a target that matches the action — control and failure modes Different bad outcomes need different controls. The branches show why alternative designs fail. Control design Align target definition with the intended action. Different bad outcomes need different controls. Failure mode 1 Combine every negative event into one label. The objective becomes confused. avoid Failure mode 2 Select the algorithm before defining success. It may optimize the wrong thing. avoid Failure mode 3 Ignore immature outcomes. Recent cases can be mislabeled as safe. avoid — Different bad outcomes need different controls. The branches show why alternative designs fail. Chapter sources · Open image

Read the confusion matrix

Precision measures the share of positive predictions that are true positives. Recall measures the share of actual positives detected. The false-positive rate uses actual negatives as its denominator. These ratios answer different questions.

In a fictional 10,000-payment cohort, 100 are confirmed fraud. A control flags 200 payments, including 80 fraud and 120 legitimate payments. Precision is 40 percent and recall is 80 percent. The false-positive rate is 120 divided by 9,900, about 1.21 percent. Count-based metrics do not describe value-weighted loss or customer harm by themselves.

A confusion matrix is meaningful only after the positive class and evaluation population are defined. If positive means confirmed fraud, a true positive is a flagged event later established as fraud under the chosen label process. A false positive is a flagged event that meets the defined negative outcome. Unresolved cases should not silently become negatives merely because a review has not finished.

Precision and recall emphasize different errors. Precision describes the proportion of flagged events that are positive; recall describes the proportion of positives that were flagged. Neither alone states the dollar loss, the review workload, or the effect on legitimate customers. Pair model metrics with action-level measures and inspect how the results change across transaction values and relevant populations.

Inside the mechanism. The confusion matrix keeps four counts visible: target events flagged, other events flagged, target events missed, and other events passed. Precision conditions on flags; recall conditions on target events; false-positive rate conditions on other events. With rare outcomes, even a small false-positive rate can dominate the queue. Report counts as well as percentages and connect them to review capacity and the actual action cost.

A concrete example. Precision, recall, and false-positive rate describe different conditional probabilities. The same threshold can create very different review demand when prevalence changes. The rule flags 986 of 50,000 labeled payments. Of those flags, 240 meet the synthetic target, giving 24.34% precision. It misses 60 target events. Under the stated cost assumptions, residual loss and operating friction total $23,574. The important result is the connection between the population, action, capacity, and outcome—not one isolated score.

When the assumption fails. A team presents high accuracy on a low-prevalence cohort as evidence of useful detection. Report all four cells, outcome definitions, and the cost and capacity implications. The following worked sequence shows the reference condition, a stress condition, and a response condition with explicit synthetic data. These are comparative assumptions, not measured causal effects.

Follow a worked case3 conditions · 36 figures

Precision, recall, and false-positive rate describe different conditional probabilities. The same threshold can create very different review demand when prevalence changes.

Confusion-matrix denominators: reference condition figure preview

01Reference conditionConfusion-matrix denominators

Confusion-matrix denominators: failure and stress figure preview

02Failure and stressConfusion-matrix denominators

Confusion-matrix denominators: controlled response figure preview

03Controlled responseConfusion-matrix denominators

Read the confusion matrix Read the confusion matrix — the distinction These concepts answer different questions. Read each definition in the context of the section. Precision 80 true positives out of 200 flags Recall 80 detected fraud out of 100 actual fraud — These concepts answer different questions. Read each definition in the context of the section. Chapter sources · Open image

Read the confusion matrix Confusion-matrix example Fictional teaching record. Fraud missed. Confusion-matrix example Illustrative data; not a real customer record or a prescribed policy. True positives 80 Fraud correctly flagged False positives 120 Legitimate payments flagged False negatives 20 Fraud missed One accuracy number can hide important errors — Fictional teaching record. Fraud missed. Chapter sources · Open image

Read the confusion matrix Read the confusion matrix — control and failure modes One accuracy number can hide important errors. The branches show why alternative designs fail. Control design Report precision recall and the relevant costs. One accuracy number can hide important errors. Failure mode 1 Call 99 percent accuracy sufficient. Rare events can make trivial models look good. avoid Failure mode 2 Divide false positives by all flags for FPR. That is not the false-positive-rate denominator. avoid Failure mode 3 Ignore payment values. Counts and financial loss differ. avoid — One accuracy number can hide important errors. The branches show why alternative designs fail. Chapter sources · Open image

Calibrate probabilities before using dollars

A calibrated probability estimate should correspond to observed outcome frequency in comparable groups, within sampling uncertainty. Ranking quality and calibration differ. A model can sort risky payments well while overstating their absolute probability.

Use held-out data appropriate to the deployment population and examine reliability across score bands and segments. If payments assigned 10 percent risk develop about 2 percent observed fraud, the score may be miscalibrated or the evaluation population may differ. Expected-loss decisions need probability meaning, not just a useful ranking. Calibration itself can drift as behavior and selection change.

Calibration compares predicted probabilities with observed frequencies in an appropriate evaluation set. Among sufficiently many events predicted near 2%, the observed positive rate should be near 2% if the probabilities are well calibrated for that setting. This does not mean any particular event will produce 2% of a loss. It also does not guarantee calibration after the traffic or policy changes. A model can rank risky events well and still produce probabilities that are unsuitable for expected-dollar calculations without further work.

Inside the mechanism. A calibrated probability has a frequency interpretation for comparable predictions in the evaluated population. Ranking quality alone does not establish that interpretation. Reliability diagrams compare predicted probabilities with observed frequencies in bins, with uncertainty and sample size considered. Fit calibration using data separated from the model’s fitting data and evaluate on appropriate held-out observations. A lower Brier score can reflect discrimination as well as calibration, so it is not a pure calibration verdict.

A concrete example. Ranking separates higher-risk cases from lower-risk cases. Calibration checks whether stated probabilities agree with observed outcome frequencies in comparable groups. The rule flags 735 of 17,300 mature decisions. Of those flags, 485 meet the synthetic target, giving 65.99% precision. It misses 121 target events. Under the stated cost assumptions, residual loss and operating friction total $50,170. The important result is the connection between the population, action, capacity, and outcome—not one isolated score.

When the assumption fails. A ranking score is multiplied by dollars as if it were a probability. Validate probability calibration on held-out data and keep severity and exposure assumptions explicit. The following worked sequence shows the reference condition, a stress condition, and a response condition with explicit synthetic data. These are comparative assumptions, not measured causal effects.

Follow a worked case3 conditions · 36 figures

Ranking separates higher-risk cases from lower-risk cases. Calibration checks whether stated probabilities agree with observed outcome frequencies in comparable groups.

Calibrated loss estimates: reference condition figure preview

01Reference conditionCalibrated loss estimates

Calibrated loss estimates: failure and stress figure preview

02Failure and stressCalibrated loss estimates

Calibrated loss estimates: controlled response figure preview

03Controlled responseCalibrated loss estimates

Calibrate probabilities before using dollars Calibrate probabilities before using dollars — the distinction These concepts answer different questions. Read each definition in the context of the section. Ranking Higher scores tend to have more events Calibration Stated probabilities match observed frequencies — These concepts answer different questions. Read each definition in the context of the section. Chapter sources · Open image

Calibrate probabilities before using dollars Calibration check Fictional teaching record. Do not price loss from the raw score. Calibration check Illustrative data; not a real customer record or a prescribed policy. Predicted band 10 percent Model estimate Observed rate 2 percent Held-out cohort result Action investigate calibration and mix Do not price loss from the raw score A ranking score is not automatically a probability — Fictional teaching record. Do not price loss from the raw score. Chapter sources · Open image

Calibrate probabilities before using dollars Calibrate probabilities before using dollars — control and failure modes A ranking score is not automatically a probability. The branches show why alternative designs fail. Control design Validate probability meaning before expected-loss use. A ranking score is not automatically a probability. Failure mode 1 Multiply any score by dollars. Its scale may have no probability interpretation. avoid Failure mode 2 Check only overall calibration. Segments can differ. avoid Failure mode 3 Calibrate on the same fitted data only. Evaluation can be overly optimistic. avoid — A ranking score is not automatically a probability. The branches show why alternative designs fail. Chapter sources · Open image

Split data by time and entity

Random splits can leak related records across training and evaluation. The same account, device cluster, or merchant can appear in both sets, making generalization look easier than it is. Temporal splits better reflect predicting later activity, but they also need entity and label-availability checks.

Define the evaluation question: future activity from known customers, new customers, new merchants, or another population. Choose splits that test it. Keep model selection separate from final evaluation. Repeatedly tuning against a holdout turns it into development data even if the filename still says test.

Inside the mechanism. Split data according to the leakage mechanism. Time separation prevents later observations from entering earlier evaluation; entity separation reduces reuse of related identities or coordinated groups across partitions. Both may be necessary. Preserve the feature availability cutoff and label maturity window. A random row split can look excellent while measuring memorization of shared entities or future information unavailable to production.

A concrete example. A random row split can put the same customer, device cluster, or future-derived feature on both sides of evaluation. Apparent performance can then reflect leakage. The comparison arm has 346/4800 adverse outcomes (7.21%) and the treatment arm has 318/4800 (6.62%). The absolute difference is -0.58 percentage points, with an illustrative large-sample 95% interval from -1.60 to 0.43. Interpretation depends on assignment integrity, outcome maturity, independence, and the actual decision being evaluated.

When the assumption fails. Transactions from one coordinated group appear in training and evaluation. Use time-aware and entity-aware evaluation with a documented boundary for shared information. The following worked sequence shows the reference condition, a stress condition, and a response condition with explicit synthetic data. These are comparative assumptions, not measured causal effects.

Follow a worked case3 conditions · 36 figures

A random row split can put the same customer, device cluster, or future-derived feature on both sides of evaluation. Apparent performance can then reflect leakage.

Time and entity leakage: reference condition figure preview

01Reference conditionTime and entity leakage

Time and entity leakage: failure and stress figure preview

02Failure and stressTime and entity leakage

Time and entity leakage: controlled response figure preview

03Controlled responseTime and entity leakage

Split data by time and entity — the distinction

Random record split: May share related entities across sets
Deployment-like split: Tests the intended future use

These concepts answer different questions. Read each definition in the context of the section. Chapter sources · Open image

Split data by time and entity Split audit Fictional teaching record. Rebuild the evaluation. Split audit Illustrative data; not a real customer record or a prescribed policy. Account appears in both sets Potential dependency Feature post-outcome record Temporal leakage Fix review split and cutoff Rebuild the evaluation Record randomness alone is not enough — Fictional teaching record. Rebuild the evaluation. Chapter sources · Open image

Split data by time and entity Split data by time and entity — control and failure modes Record randomness alone is not enough. The branches show why alternative designs fail. Control design Design splits for the actual generalization claim. Record randomness alone is not enough. Failure mode 1 Assume a random seed prevents leakage. It only makes the split repeatable. avoid Failure mode 2 Tune repeatedly on the final holdout. It stops being independent evidence. avoid Failure mode 3 Ignore delayed labels. Future outcomes can leak into training. avoid — Record randomness alone is not enough. The branches show why alternative designs fail. Chapter sources · Open image

Monitor the model as part of a system

Track input quality, score distribution, calibration, ranking, latency, coverage, and mature outcomes. Data drift means inputs changed; performance drift means outcome quality changed. One can occur without the other.

Define action thresholds with owners and context. A score-distribution shift may come from a new merchant mix rather than a model defect. A stable score distribution can still hide deteriorating calibration. Monitor the decision policy and human overrides too. Customers experience the complete system, not an isolated model artifact.

Inside the mechanism. Monitor the decision system around the model: input quality, score distribution, action delivery, queue capacity, overrides, mature outcomes, and relevant population changes. A stable score distribution does not prove stable performance. Labels can stop arriving while scores remain normal. Tie alerts to a supported response and distinguish data faults, population drift, and model degradation before choosing a remedy.

A concrete example. A model can retain its score distribution while its inputs, labels, or downstream actions fail. Monitoring should join technical quality with mature business outcomes. The daily source population is 39,000 items, but 780 are outside the completed monitoring run. The included population creates 459 hits and 376 unique cases. With 95 cases already open and capacity for 460, the queue closes at 11. Coverage, duplicate work, and staffing are separate causes; reducing one number does not prove that the overall control improved.

When the assumption fails. The model dashboard stays green while review capacity drops and label delivery stops. Monitor input coverage, action delivery, queue age, and outcome maturity together. The following worked sequence shows the reference condition, a stress condition, and a response condition with explicit synthetic data. These are comparative assumptions, not measured causal effects.

Follow a worked case3 conditions · 36 figures

A model can retain its score distribution while its inputs, labels, or downstream actions fail. Monitoring should join technical quality with mature business outcomes.