Sigma

Sigma gives you a quick, fully automatic, statistically robust look at a dataset. It turns raw tabular data into a readable, statistically grounded binary tree. It is not designed to compete with state-of-the-art machine learning models on raw predictive performance. Those models are great when you need the best predictions, but they require effort to use - analyzing their behavior and understanding their main decision drivers are time-consuming. So Sigma is a quick, robust, straightforward choice for reliable diagnostics on new datasets, one-off segmentation studies, and model drift investigations, and it helps you examine the outputs of more complex models before trusting those models.

scikit-learn's decision trees tend to be biased in their variable selection, require manual tuning of depth and minimum-samples thresholds to stay readable, and routinely grow into hairy, hard-to-interpret structures. They are biased in the features they select, and they quickly overfit noise because their rule for splitting a node further is not grounded in statistics. Sigma uses permutation-based conditional-inference splits instead: at every node it tests each candidate variable for a real association with the response, picks the strongest, and stops the moment none is statistically significant. The tree finds its own depth automatically: no manual cutoff, no rule of thumb on samples per leaf. The single knob is the false-positive rate of each split test (the alpha level, 5% by default); raise it for a deeper exploratory tree, lower it for a stricter one. The result is a tree you can trust as a diagnostic.

How Sigma compares to the common alternatives:

Yes, and it is great for quickly going to production in constrained environments, e.g., to deploy a model in a few lines of code or SQL. But I have something better. Fitting a Sigma tree gives you a robust first-approximation predictive model with very little effort: no tuning, no feature engineering, no encoding of categoricals, and a tree you can read end to end. It is a sound baseline you can deploy and trust. The tradeoff is subtlety for simplicity. A Sigma tree captures the dominant structure in the data but misses the finer interactions and gradients a richer model could exploit. When you need that accuracy without giving up interpretability, reach for Tau, my breakthrough engine for tabular data. It turns weak signals into outstanding predictions.

Yes. ClassificationTree, RegressionTree, SurvivalTree, and RankingTree implement the standard scikit-learn estimator API, so they slot into Pipeline, ColumnTransformer, GridSearchCV, cross-validation utilities, and model serialization without any custom glue. Pandas DataFrames are first-class inputs: pass them directly to fit and predict, declare categorical columns by name, and keep your column metadata throughout training and inference. NumPy arrays are of course also supported. Categorical features are handled natively (no upstream one-hot encoding needed), and sample weights are fully supported through all the statistical significance tests.

Sigma supports the classic confidence intervals; the right one depends on the coverage guarantees you want. For classification: Clopper-Pearson, Jeffreys, and Wilson. For regression: Bayesian bootstrap, BCa, normal, Student-t, log-normal, gamma, Poisson, exponential, and beta. For survival outcomes: Brookmeyer-Crowley for the median survival time, log-log Greenwood for the survival probability at a fixed time, and Klein-Moeschberger integrated Greenwood for the restricted mean survival time (RMST). You pick the one that best fits your target's distribution.

On small datasets, the statistical stopping rule and the confidence intervals on leaves keep the tree honest about what it does and does not know. On larger datasets, fit time scales linearly with the number of observations and with the number of candidate variables, so Sigma's runtimes stay predictable as the data grows.

Python 3.10 and above, on every platform supported by NumPy and SciPy. Sigma depends only on NumPy, SciPy, and scikit-learn. Graphviz rendering is available as an optional extra for visualization, but it is not required to fit or use models.

A single pip install ars-sigma command installs the library. There are no compiled extensions to build, which makes installation simple. Everything else (visualization examples, deployment notes, source, and issue tracker) is on the GitHub repository.

Sigma is source-available under the Sigma License (not open source). You can use and redistribute it, keeping the license file with every copy, but not feed it to AI systems, except to generate your own client code that calls its public API. Modifications are accepted as upstream Contributions to the canonical repository. The license is revocable. A non-revocable license is available through the contact page. Contributions, issues, and feedback are welcome on the GitHub repository.

Reach out through the contact page. Your message goes directly to me.