inferust

Statistical modeling for Rust — a statsmodels-inspired library.

inferust fills the gap between Python's statsmodels / scipy.stats and the Rust ecosystem. It gives you regression summaries, hypothesis tests, descriptive stats, and correlation matrices with the same depth of output you'd expect from Python — p-values, confidence intervals, AIC/BIC, significance stars, and all.

Features

Installation

Add to your Cargo.toml:

[dependencies]
inferust = "0.1"

Quick start

OLS Regression

use inferust::regression::Ols;

let x = vec![
    vec![2.0, 3.1],
    vec![5.0, 3.7],
    vec![8.0, 3.5],
    vec![11.0, 3.6],
];
let y = vec![55.0, 70.0, 80.0, 90.0];

let result = Ols::new()
    .with_feature_names(vec!["hours_studied".into(), "prior_gpa".into()])
    .fit(&x, &y)
    .unwrap();

result.print_summary();

Output:

═══════════════════════════════════════════════════════════════════
                     OLS Regression Results
═══════════════════════════════════════════════════════════════════
 Dep. variable: y          Observations  : 4
 R²           : 0.998102   Adj. R²       : 0.994305
 F-statistic  : 262.7732   F p-value     : 0.039405
 AIC          : 14.7316    BIC           : 12.0167
───────────────────────────────────────────────────────────────────
Variable               Coef       Std Err         t      P>|t|
───────────────────────────────────────────────────────────────────
const              -5.654762    5.033740    -1.1234   0.460565
hours_studied       4.130952    0.177951    23.2141   0.027430  *
prior_gpa           8.166667    1.490421     5.4793   0.115581
───────────────────────────────────────────────────────────────────
 Significance codes:  *** p<0.001  ** p<0.01  * p<0.05  . p<0.1
═══════════════════════════════════════════════════════════════════

Hypothesis tests

use inferust::hypothesis::{ttest, anova, chisq};

// Paired t-test
let before = vec![72.0, 68.0, 75.0, 80.0, 65.0];
let after  = vec![78.0, 74.0, 80.0, 85.0, 72.0];
ttest::paired(&before, &after).unwrap().print();

// Two-sample Welch t-test
ttest::two_sample(&group_a, &group_b).unwrap().print();

// One-way ANOVA
anova::one_way(&[&group1, &group2, &group3]).unwrap().print();

// Chi-squared goodness-of-fit
chisq::goodness_of_fit(&observed, None).unwrap().print();

// Chi-squared test of independence
chisq::independence(&contingency_table).unwrap().print();

Descriptive statistics

use inferust::descriptive::Summary;

let data = vec![4.2, 7.8, 5.1, 9.3, 3.6, 8.4];
Summary::new(&data).unwrap().print();
// ─────────────────────────────
//  n          : 6
//  mean       : 6.400000
//  std        : 2.282176
//  min        : 3.600000
//  25%        : 4.575000
//  50%        : 6.150000
//  75%        : 8.250000
//  max        : 9.300000
//  skewness   : -0.058732
//  kurtosis   : -1.504070
// ─────────────────────────────

Correlation

use inferust::correlation;

let r = correlation::pearson(&x, &y).unwrap();
let rs = correlation::spearman(&x, &y).unwrap();

let matrix = correlation::correlation_matrix(&[hours, gpa, scores]).unwrap();
correlation::print_correlation_matrix(&matrix, &["hours", "gpa", "scores"]);

OLS builder options

Ols::new()                                        // intercept on by default
    .with_feature_names(vec!["x1".into()])        // label columns
    .no_intercept()                               // force through origin
    .fit(&x, &y)
    .unwrap();

OlsResult also exposes .predict(&x) for out-of-sample predictions and all raw fields (coefficients, residuals, r_squared, p_values, etc.) for programmatic use.

Error handling

All fallible functions return inferust::Result<T> (an alias for Result<T, InferustError>):

use inferust::InferustError;

match result {
    Err(InferustError::SingularMatrix)           => { /* perfect multicollinearity */ }
    Err(InferustError::InsufficientData { .. })  => { /* too few rows */ }
    Err(InferustError::DimensionMismatch { .. }) => { /* X rows ≠ y length */ }
    Err(InferustError::InvalidInput(msg))        => { /* other input problem */ }
    Ok(r) => { /* use result */ }
}

Dependencies

Roadmap

• Logistic regression (GLM with logit link)
• Ridge / Lasso regularization
• Durbin-Watson and Breusch-Pagan diagnostic tests
• Tukey HSD post-hoc test (after ANOVA)
• Time-series: ARIMA / ACF / PACF
• Weighted OLS

Contributions welcome — open an issue or PR!

License

MIT — see LICENSE.