Fraud is an adversarial, extremely imbalanced problem with delayed and noisy labels and asymmetric costs. The patterns that make fraud models look great offline and fail live are well known:
- features computed with future information
- random splits
- accuracy or ROC AUC as the headline
- thresholds picked without the alert budget or the money involved
This skill builds the features, the protocol and the evaluation around those realities:
scripts/fraud_velocity_features.pycomputes leak-free velocity featuresscripts/gen_fraud_model_script.pywrites an out-of-time LightGBM + anomaly-score modelscripts/fraud_eval.pyevaluates a score in alerts, money and cost
Trigger Conditions
Activate this skill when the user:
- Builds or reviews a fraud, scam, account-takeover, promo-abuse, chargeback or AML model
- Needs velocity / behavioural features for transactions or events
- Must set a fraud threshold, size an alert queue, or estimate money saved
- Sees a fraud model that degrades after launch or looks too good offline
General imbalance techniques and threshold tuning: imbalanced-class-handler. Unsupervised detectors: anomaly-detection-model-builder. Leakage: target-leakage-detector.
Workflow
1. LABELS -> define fraud (confirmed chargeback, analyst decision, customer report), label delay, exclusions
2. FEATURES -> python scripts/fraud_velocity_features.py tx.csv --entity card_id --categorical merchant_category country -o feats.csv
3. MODEL -> python scripts/gen_fraud_model_script.py --data tx.csv --entity card_id --categorical merchant_category country -o model.py
4. EVALUATE -> python model.py; python scripts/fraud_eval.py test_scored.csv --alerts-per-day 50
5. OPERATE -> threshold from cost and capacity; rules + model; monitoring of AP by week and fraud mixLabels and Splits
| Issue | Rule |
|---|---|
| Label delay | chargebacks arrive weeks to months later. Leave a gap between train end and test start (--gap-days), and do not train on immature recent labels |
| Label noise | friendly fraud, unconfirmed alerts and "no label = legit" bias. Document the label source |
| Selection bias | blocked transactions never get labels. Keep a small random hold-out that is not blocked, or use rejected-inference caveats |
| Splits | out of time only. Random splits let the model learn fraud rings it would not have seen yet |
| Entity leakage | features or splits that let the same card or ring appear in train and test with future knowledge |
Features That Work
| Family | Examples | Leak-free rule |
|---|---|---|
| Velocity | transactions and amount per card in 1h / 24h / 7d | windows end before the current event (closed="left") |
| Behaviour deviation | amount / card's past mean, z-score, unusual hour | statistics from past events only |
| Novelty | first time this merchant category / country / device / IP for the entity | the "seen" set is updated after scoring |
| Network | shared devices, emails, IPs, cards per device; graph features | as of the event time |
| Merchant / context | merchant risk tier, MCC, channel, card-not-present | known at authorisation |
| Unsupervised | Isolation Forest / autoencoder score fitted on the training period | fit on train only |
fraud_velocity_features.py (stdlib, streaming) and the generated script's pandas features produce
identical values, and both match pandas rolling(closed="left"). The same definitions can therefore be
served online. Feature catalogue and real-time computation: fraud-features.md.
Evaluating a Fraud Score
| Question | Metric | fraud_eval.py |
|---|---|---|
| Does it rank fraud first? | average precision (baseline = fraud rate); ROC AUC only as secondary | ranking |
| What does the alert queue look like? | precision, recall and alerts/day at the top 0.5-5% | alert rates |
| How much money is stopped? | value detection rate (fraud amount caught / total) | every table |
| How many good customers are hit? | FPR, recall at FPR 0.1% / 1% | at FPR |
| Where should the threshold be? | minimal expected cost: missed amount x loss rate + review cost + friction | cost |
| Can the team handle it? | best threshold within the alerts-per-day capacity | --alerts-per-day |
| Is it stable? | AP and fraud rate by week | by week |
On the fixture: AP 0.83 against a random 0.023. The cost-optimal threshold reviews 121 alerts, catches 98.6% of the fraud value and costs 705 against 11,807 without a model. A capacity of one alert per day lowers value detection to 73%. Metrics match scikit-learn (ROC AUC, AP). Details and the cost model: fraud-evaluation.md.
Production Design
- Rules + model: hard rules for known patterns and regulatory blocks, the model for ranking, and analyst review for the uncertain band.
- Decisions: approve / step-up authentication / review / decline. Each has its own threshold and cost.
- Real time: sub-100 ms scoring with online feature stores (velocity counters in Redis or Flink).
- Feedback loop: blocked fraud gets no chargeback label. Record analyst outcomes and keep a random control.
- Adversarial drift: retrain frequently, monitor AP by week, the feature distributions and new fraud-pattern clusters.
Operations, monitoring and governance: fraud-operations.md.
Output Format
When designing or reviewing a fraud model, deliver:
- Label definition - source, delay, exclusions, the gap used.
- Features - families, point-in-time guarantees, online computation plan.
- Model and split - out-of-time design, model, importance, the anomaly component.
- Operating point - the alert-rate table, value detection, the cost-optimal and capacity thresholds.
- Operations - decisions per band, monitoring (AP by week, drift), the retraining cadence.
Scripts
python scripts/fraud_velocity_features.py transactions.csv --entity card_id --time-col timestamp --amount-col amount \
--categorical merchant_category country -o features.csvExit codes: 0 written, 2 usage or read error.
python scripts/gen_fraud_model_script.py --data transactions.csv --entity card_id --label-col is_fraud -o fraud_model.py
python scripts/gen_fraud_model_script.py --data tx.csv --entity account_id --categorical merchant_category country --gap-days 14Exit codes: 0 written, 2 usage error.
python scripts/fraud_eval.py transactions.csv --label-col is_fraud --score-col score --amount-col amount
python scripts/fraud_eval.py scored.csv --alerts-per-day 20 --review-cost 5 --friction-cost 10 --loss-rate 0.8Exit codes: 0 evaluated, 2 usage or read error.
Additional Resources
| File | Content |
|---|---|
| fraud-features.md | feature families, point-in-time computation, online counters, graph signals |
| fraud-evaluation.md | metrics under imbalance, value detection, cost model, capacity, label delay in evaluation |
| fraud-operations.md | rules + model, decision bands, feedback loops, drift and adversaries, monitoring, governance |
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.