stepzero

Task-first ML baselines. Run the simplest thing that could work.

Before reaching for XGBoost or a neural net, run stepzero. It fits the simplest sensible model for your task, compares a few alternatives, and tells you whether your baseline is good enough or what to try next.

import stepzero as sz

result = sz.classification(X, y)
print(result)
# ClassificationResult(best='logistic', accuracy=0.960, headroom='low')

print(result.headroom)
# [low] Score of 0.96 with low variance (±0.012). The simple baseline is already
# performing well. Trying a gradient boosted tree is unlikely to offer a meaningful improvement.

---

Install

pip install stepzero

Requirements: Python 3.10+, numpy, pandas, scikit-learn, scipy.

---

Tasks

Classification

result = sz.classification(X, y)

result.best_model          # fitted sklearn Pipeline — call .predict(X_new) directly
result.best_model_name     # "logistic" | "tree" | "naive_bayes"
result.scores              # [ModelScore(name, score, metric), ...]
result.feature_importance  # pd.Series sorted by importance
result.headroom            # HeadroomSignal(level, reason)

• Methods: logistic regression, decision tree, naive bayes
• Metric: accuracy (5-fold stratified CV)

Regression

result = sz.regression(X, y)

result.best_model_name     # "ridge" | "tree"
result.feature_importance  # normalized importances as pd.Series
result.headroom

• Methods: ridge, decision tree
• Metric: RMSE (5-fold CV)

Forecasting

result = sz.forecasting(series, horizon=12)

result.forecast        # pd.Series with future timestamps as index
result.best_model_name # "seasonal_naive" | "linear_trend"
result.scores          # MAE per model
result.headroom

• Methods: seasonal naive, linear trend
• Parameters: horizon, freq (optional — inferred from DatetimeIndex), cv_splits
• Metric: MAE (time-series CV)

Anomaly Detection

result = sz.anomaly_detection(series)

result.anomalies   # pd.Series[bool], same index as input
result.scores      # raw anomaly scores
result.method      # "zscore" | "iqr"
result.threshold   # auto-determined threshold
result.headroom

• Methods: z-score, IQR
• Parameters: threshold (optional — auto-set to flag ~5% of points), method
• Metric: inter-method agreement

Text Classification

result = sz.text_classification(texts, labels)

result.best_model_name        # "tfidf_logistic" | "tfidf_naive_bayes"
result.top_features_per_class # {"class_0": ["word1", ...], ...}
result.headroom

• Methods: TF-IDF + logistic regression, TF-IDF + naive bayes
• Metric: accuracy (5-fold stratified CV)

Clustering

result = sz.clustering(X, k_range=(2, 10))

result.best_k    # selected number of clusters
result.labels    # cluster assignment per sample (np.ndarray)
result.centers   # cluster centroids in original feature space
result.scores    # silhouette score per k tried
result.headroom

• Methods: k-means
• Parameters: k_range
• Metric: silhouette score

---

The headroom signal

Every result has a .headroom attribute:

result.headroom.level   # "low" | "medium" | "high"
result.headroom.reason  # actionable explanation + what to try next
print(result.headroom)
# [medium] CV accuracy of 0.81 ± 0.04. A 19% gap to ceiling remains.
# A gradient boosted tree (e.g., XGBoost or LightGBM) is a natural next step.

• low means that the simple model is already doing well; complexity buys little
• medium means that meaningful headroom remains; a tuned model may help
• high means that the baseline is underperforming; a more complex model is likely worth it

---

Design philosophy

• Task-first, not model-first. You describe the problem; stepzero picks the approach.
• Opinionated defaults. Auto-scaling for linear models, missing value imputation, sensible eval.
• No false modesty. The models are genuinely simple — logistic regression, decision trees, seasonal naive. No AutoML hidden underneath.
• Ready to deploy. result.best_model is a fitted sklearn Pipeline. Call .predict() on new data immediately.
• Minimal footprint. Only numpy, pandas, scikit-learn, and scipy. No optional heavy dependencies required for core functionality.

---

When to use stepzero

• ✅ Starting a new ML project and want a defensible baseline in 5 minutes
• ✅ Proving (or disproving) that a simple model is good enough
• ✅ Teaching or demonstrating ML without the XGBoost-first bias
• ✅ Kaggle competitions — establish your baseline before tuning

---

Contributing

Contributions are welcome. Please read CONTRIBUTING.md for the workflow.

In short: branch from develop, open a PR targeting develop. All PRs run the test suite automatically across Python 3.10–3.12.

Reporting issues

Open an issue on GitHub. Include your Python version, stepzero version, and a minimal reproducible example.

---

License

MIT