The chain
KPI -> decision metric -> offline proxy -> online confirmation| KPI | Decision | Offline metric | Online metric |
|---|---|---|---|
| Monthly churn rate | who gets a retention call (400/week) | recall@400 of 30-day churners (time holdout) | saves per 400 calls vs holdout group |
| Fraud losses | block / step-up at authorisation | value-weighted recall at 0.2% FPR | fraud $ per $1M volume; false-decline rate |
| Gross margin from promotions | who gets a coupon | Qini / uplift@K | incremental margin in an A/B test |
| Revenue from sales pipeline | rep queue order | opportunities in top 750 leads/day | pipeline $ per rep-hour |
| Stock-outs and waste | order quantity | pinball loss at the service quantile, bias | stock-out rate, waste rate |
| Support resolution time | ticket routing | top-1 accuracy, reassignment rate | median time to resolution |
| Search conversion | ranking | NDCG@10 on judged queries | conversion per search session |
| Credit losses | approve / limit | AUC + calibration by score band | default rate per band, approval rate |
Cost matrix workshop
Fill with the business owner; use ranges when unsure and test the threshold's sensitivity.
| Predicted no action | Predicted action | |
|---|---|---|
| Event would not happen | 0 | C_fp: action cost + side effects (annoyance, discount given away) |
| Event would happen | C_fn: loss not prevented | C_tp: action cost − value saved x success rate |
Example (retention): call cost $8; annoyance cost $2; saved-customer value $600 x 30% save rate = $180.
C_fp = $10, C_tp = $8 − $180 = −$172, C_fn = $0 relative to doing nothing (the loss happens either way).
Act when p·(C_fn − C_tp) > (1 − p)·C_fp -> p·172 > (1 − p)·10 -> p > 0.055.
Threshold policies
| Policy | When | Implementation |
|---|---|---|
| Top-K per period | fixed capacity | rank, take K; re-evaluate K when capacity changes |
| Cost-optimal | costs known, capacity flexible | calibrated p > C_fp / (C_fp + C_fn − C_tp) |
| Constraint | regulation or SLA ("≥ 95% recall") | highest threshold meeting it on validation |
| Bands | human tiers (auto-approve / review / decline) | two thresholds; measure each band's precision |
| Per-segment | different costs per segment | segment-specific thresholds; check fairness rules first |
Recheck the threshold after retraining: a new model's score distribution shifts even at the same AUC.
Guardrail catalogue
| Area | Guardrail |
|---|---|
| Fairness | TPR/FPR ratio between groups within 0.8-1.25; per-group calibration |
| Customer | complaint rate, unsubscribe rate, false-decline rate |
| Operations | alerts per analyst per day, queue backlog |
| Business | margin, approval rate, average order value |
| Model health | PSI of scores and top features < 0.2, metric by week, latency p99 |
Offline/online disagreement
| Symptom | Likely cause | Fix |
|---|---|---|
| Offline better, online flat | the action does not change outcomes for the selected cases | uplift framing, better action |
| Offline better, online worse | leakage or train/serve skew | check features at serving time; shadow mode |
| Online effect shrinks over weeks | feedback loop or novelty effect | holdout groups, longer test |
| Offline metric noisy | small test set | CI, longer holdout window |
Writing the success criteria
Primary: recall@400 per week of customers who cancel within 30 days, on the Apr-Jun 2024 time holdout.
Target >= 0.45 (current rule 0.30). Threshold: top 400 by score each Monday.
Secondary: PR-AUC, ROC-AUC, lift@400.
Guardrails: complaint rate of called customers <= 1.5%; recall@400 for each region >= 0.35;
score PSI vs training < 0.2.
Online: saves per 400 calls vs a 5% untreated random holdout over 6 weeks; launch if +5 saves/week with
95% CI excluding 0.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.