Struct VarModel 

pub struct VarModel { /* private fields */ }

After fitting, the model verifies that the VAR companion matrix has a spectral radius below stability_threshold (default: 1.05). Models exceeding this are rejected as unstable. The QR-based eigenvalue computation correctly captures complex conjugate eigenvalue pairs (oscillatory dynamics) that power iteration would miss. Callers can override the threshold via new_with_stability_threshold().

Implementations

impl VarModel

pub fn new(lag_order: Option<usize>) -> Self

Creates a VAR model with the given lag order (default: 1).

pub fn new_with_stability_threshold( lag_order: Option<usize>, stability_threshold: Option<f64>, ) -> Self

Creates a VAR model with a custom stability threshold.

The stability_threshold controls the maximum allowed spectral radius of the VAR companion matrix (default: 1.05). Set to a value > 1.0 to tolerate near-unit-root processes, or closer to 1.0 for stricter stability enforcement.

impl VarModel

pub fn granger_causality( &self, ctx: &MultiSeriesContext, ) -> Result<Vec<GrangerCausalityResult>, MultivariateError>

Perform pairwise Granger causality tests for all series pairs.

For each pair (cause → effect), fits an unrestricted VAR equation (with all series’ lags) and a restricted equation (excluding the cause’s lags), then compares RSS via an F-test.

Requires the model to be fitted first via fit().

F-test

$$F = \frac{(RSS_r - RSS_u) / p}{RSS_u / (T - k_{full})}$$

where $p$ = lag order (number of excluded regressors), $T$ = number of usable observations, and $k_{full}$ = number of regressors in the unrestricted equation (1 + p × k).

pub fn granger_causality_robust( &self, ctx: &MultiSeriesContext, ) -> Result<Vec<GrangerCausalityResult>, MultivariateError>

HC3-robust pairwise Granger causality tests.

Same hypothesis as granger_causality but uses the HC3 (MacKinnon & White 1985) covariance estimator, which is valid under arbitrary conditional heteroscedasticity.

For each pair (cause → effect) the unrestricted OLS equation is fitted, residuals are obtained, and a Wald test with HC3 sandwich covariance is computed:

$$W = (R\hat\beta)^\top (R,\hat V_{HC3},R^\top)^{-1}(R\hat\beta)$$

where $R$ selects the $p$ coefficients of the candidate cause. $F = W / p$ is compared against $F(p, T - kp - 1)$.

Trait Implementations

impl MultivariateForecastModel for VarModel

fn fit( &mut self, ctx: &MultiSeriesContext, target_idx: usize, ) -> Result<(), MultivariateError>

Fit the model.

fn predict( &self, horizon: usize, ) -> Result<MultivariateForecastResult, MultivariateError>

Predict horizon steps ahead.

fn supports_online_learning(&self) -> bool

Whether the model supports incremental online learning.