Rolling-origin (time series cross-validation)
history ----------------------------|origin 1|--h-->
history --------------------------------------|origin 2|--h-->
history ----------------------------------------------------|origin 3|--h-->- Expanding window (all data before the origin) is the default. A sliding window suits regime changes.
- Space origins by the horizon (non-overlapping test windows) or by one season.
- Refit at every origin, or at least re-estimate state. Never use information after the origin.
- Include the periods that matter: if December decides the business, backtest Decembers.
Metrics
| Metric | Formula | Use | Watch out |
|---|---|---|---|
| MAE | mean |y - f| | units, robust | favours the median: forecasts of 0 for intermittent series |
| RMSE | sqrt(mean e^2) | penalises large misses | scale-dependent |
| MASE | MAE / in-sample MAE of the seasonal naive | compare across series; < 1 beats the one-step seasonal naive | multi-step MASE > 1 is common |
| RMSSE | sqrt(mean e^2 / in-sample MSE of the seasonal naive) | intermittent demand (M5) | |
| WAPE | sum |e| / sum |y| | business reporting, aggregates | dominated by large series |
| Bias | sum e / sum y | systematic over- or under-forecasting | report it always |
| MAPE | mean |e / y| | avoid | undefined with zeros, asymmetric |
| sMAPE | mean 2|e| / (|y| + |f|) | M3/M4 comparisons only | asymmetric, unstable near zero |
| Pinball loss | quantile loss at level q | quantile forecasts | |
| Coverage / interval width | share of actuals inside the intervals | prediction intervals | compare with the nominal level |
| CRPS | integrated pinball | full predictive distributions |
Skill score: 1 - metric(model) / metric(baseline). The baseline is the seasonal naive unless a stronger
simple method exists (it often does: Holt-Winters beat snaive by 30% on the fixture).
Error by horizon
Errors grow with the horizon. Report at least h=1, h/2 and h. A model that wins at h=1 but loses at h=28 is
the wrong model for a 28-day planning process. ts_backtest.py prints the MAE by horizon step for the best
model.
Pitfalls
| Pitfall | Effect | Fix |
|---|---|---|
| single train/test split | lucky or unlucky period decides | rolling origin |
| tuning on the test origins | optimistic results | nested: tune on earlier origins, evaluate on later ones |
| comparing on different origins or horizons | apples vs oranges | one protocol for every model (--forecasts) |
| filling gaps with future values | leakage | forward-fill or interpolate within the training window only |
| evaluating on aggregates only | hides SKU-level failure | report both levels |
| ignoring bias | systematic stock-outs or overstock | bias column and a bias alert in monitoring |
| MAPE on intermittent demand | infinite or meaningless | RMSSE, MASE, or inventory-based metrics (fill rate) |
Reporting template
| Model | MASE | RMSSE | Skill vs snaive | WAPE | Bias | MAE h1 / h14 / h28 | 80% coverage |
|---|
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.