Why not accuracy or ROC AUC alone
At 0.5% fraud, predicting "never fraud" is 99.5% accurate. ROC AUC is dominated by the huge negative class: 0.98 can still mean most alerts are false. Average precision (area under the PR curve) and metrics at the operating point reflect what analysts and customers experience. The random-baseline AP equals the fraud rate.
Operating-point metrics
| Metric | Definition | Audience |
|---|---|---|
| Precision (hit rate) | true frauds / alerts | analysts: share of useful work |
| Recall (detection rate) | frauds caught / all frauds | risk teams |
| Value detection rate | fraud amount caught / total fraud amount | finance: the money metric |
| FPR / false-positive ratio | good transactions alerted / good transactions, or false : true alerts | customer experience |
| Alerts per day | volume at the threshold | operations capacity |
Report them at business-relevant cut-offs: the top 0.5-5% of scores, FPR 0.1% / 1%, or the capacity limit.
Cost model (fraud_eval.py)
cost(threshold) = sum over missed frauds of amount x loss_rate
+ number of alerts x review_cost
+ number of false alerts x friction_costloss_rate: share of the fraud amount actually lost (recoveries and liability shift lower it).review_cost: analyst minutes x cost per minute, or the cost of step-up authentication.friction_cost: lost sales, support calls and churn from declining or challenging good customers. It is often underestimated, so ask the business. Compare with "no model" (lose all fraud) and "review everything". Report the savings and the sensitivity of the optimal threshold to the cost assumptions (rerun with different--friction-cost).
Capacity
Analyst teams handle a fixed number of alerts per day. --alerts-per-day finds the cheapest threshold within
the capacity. When the cost-optimal threshold exceeds it, quantify the money left on the table. That is the
case for more analysts, or for automating the clear cases.
Label delay in evaluation
- Evaluate only on periods whose labels have matured (for example, 60-90 days after the transactions for chargebacks).
- Recent weeks show artificially low fraud rates and recall.
fraud_eval.pyprints the fraud rate by week, and a sudden drop near the end usually means immature labels, not success.
Stability
Average precision per week varies a lot when frauds per week are few (it ranged 0.44-0.97 on the fixture). Use rolling 4-week windows or confidence intervals (bootstrap over days) before calling a degradation.
Comparing models
Use the same out-of-time test period and the same label maturity. Compare AP, the value detection rate at the same alert volume, and the cost at each model's own optimal threshold. Use paired bootstrap over days for significance.
anomaly-detection-model-builder
Build and evaluate anomaly detection for tabular, sensor and transaction data with few or no labels: define the anomaly type (point, contextual, collective), choose detectors (robust z-score/MAD and IQR rules, Mahalanobis with robust covariance, k-NN distance, LOF, Isolation Forest, one-class SVM, PCA reconstruction / autoencoders, forecasting residuals for time series), combine them by rank, explain each alert by its deviating features, set thresholds from an alert budget or the score distribution, and evaluate with precision@k, PR-AUC and recall on known incidents, or detector agreement when no labels exist. Includes pure-Python detectors (robust z, Mahalanobis, kNN, Isolation Forest, ensemble) and an evaluation script.
catboost-model-builder
Build CatBoost models that use its strengths correctly: raw categorical columns as cat_features (ordered target statistics, no manual encoding), text_features, Pool construction with NaN-safe categories, ordered boosting and has_time for temporal data, symmetric (oblivious) trees, overfitting detector with eval_set and use_best_model, losses for classification, regression (Quantile, Poisson, Tweedie) and ranking (YetiRank), class weights, monotone constraints, feature importance (PredictionValuesChange, LossFunctionChange) and SHAP values, GPU training and model export. Plans cat/text/numeric columns from a CSV and generates a runnable training script, and checks parameter sets for the conflicts CatBoost rejects (iterations vs n_estimators, Bayesian bootstrap with subsample, multiple class-weight options, depth > 16, use_best_model without eval_set).
categorical-encoding-selector
Choose and implement the right encoding for every categorical column given the model family and the data: one-hot (with infrequent-level grouping and handle_unknown), ordinal with an explicit order, cross-fitted target encoding (scikit-learn TargetEncoder, CatBoost ordered statistics, category_encoders inside CV), frequency/count encoding, feature hashing, native categorical support in LightGBM, HistGradientBoosting, XGBoost and CatBoost, prefix levels for codes (zip3, NAICS-2), and embeddings. Measures cardinality, rare-level share, unseen-level rate on a time-ordered holdout and target signal per column, emits the ColumnTransformer, and quantifies target-encoding leakage (in-sample vs out-of-fold vs holdout AUC).
classification-model-selector
Choose the right classification algorithm for tabular, text or mixed data and prove the choice with a fair bake-off: logistic regression (L2, L1, elastic net), linear SVM, naive Bayes, k-NN, RBF SVM, decision trees, random forest, extra trees, HistGradientBoosting, LightGBM, XGBoost, CatBoost, explainable boosting machines (EBM/GAM), points scorecards and MLPs, matched to data size, width, categorical and high-cardinality share, missingness, class imbalance, text, and constraints such as interpretability, monotonic constraints, calibrated probabilities, latency budgets and tiny deployable models. Ranks candidates with starting hyperparameters and cautions, then generates a bake-off script running every candidate on identical CV folds with per-fold metrics, fit time and prediction latency.
clustering-algorithm-selector
Choose the right clustering algorithm and check that clustering makes sense at all: cluster tendency (Hopkins statistic with a correlated-features caveat, nearest-neighbour isolation), then k-means / MiniBatchKMeans, Gaussian mixtures with BIC, agglomerative (Ward, average, complete, single) with dendrograms, DBSCAN, HDBSCAN, OPTICS, spectral clustering, BIRCH, mean shift, k-prototypes and k-modes for mixed or categorical data, and Gower distance, matched to data size, expected shapes, noise, varying density, known k, soft memberships, hierarchy and streaming. Covers preprocessing (scaling, encoding, dimensionality, distance metrics), validation (internal indices, bootstrap stability, external labels, business profiling) and assigning new points.
collaborative-filtering-implementer
Implement collaborative filtering recommenders that beat popularity: item-based kNN with cosine, Jaccard or asymmetric similarity, shrinkage, neighbourhood size, positive-only or mean-centred rating signals and 'because you liked X' explanations; implicit ALS with confidence weighting, BPR pairwise ranking and the closed-form EASE model in plain NumPy, tuned on a temporal validation split by recall@k and NDCG@k and refit on all data; plus user-kNN, explicit matrix factorisation, the implicit/LightFM libraries, hyperparameter ranges, cold-start fallbacks, scaling with sparse matrices and approximate nearest neighbours. Produces top-k lists in the format of the recommendation evaluator.