GeneratedAssembled automatically from moduleDocs.ts. No human has reviewed this page.

sb_bagging

Concept Overview

Bagging in which the resampling respects label overlap. The standard bootstrap assumes IID draws; with triple-barrier labels whose spans overlap, an IID bag is full of near-duplicates, the base learners end up correlated, and the variance reduction bagging promises never materialises. Sequential bootstrap instead draws each index with probability proportional to its average uniqueness given what has already been drawn, so each bag is as close to independent as the data permits.

When to Use

Use it in place of ordinary bagging whenever the labels come from labeling — that is, whenever observations overlap in time. Measure the benefit rather than assuming it: ensemble_methods::average_pairwise_prediction_correlation will tell you whether the base learners actually decorrelated, and if rho is still high the extra sampling cost bought nothing. Note that new() takes the random seed, not the ensemble size: n_estimators defaults to 10 and must be set explicitly.

Mathematical Foundations

Bagging Predictor

$\hat f(x)=\frac{1}{B}\sum_{b=1}^{B} f_b(x)$

where $B$ = n_estimators and $f_b$ is the base learner fitted to the $b$ -th resample. Note that $B$ defaults to $10$ , not to the constructor argument, which is the random seed.

Sequential Bootstrap Draw

$\Pr\!\left[i\mid\varphi\right]=\frac{\bar u_i(\varphi)}{\sum_j \bar u_j(\varphi)},\qquad \bar u_i(\varphi)=\frac{1}{|T_i|}\sum_{t\in T_i}\frac{1}{1+c_t(\varphi)}$

where $\varphi$ is the set of indices drawn so far, $T_i$ the bars spanned by observation $i$ ‘s label, and $c_t(\varphi)$ the number of already-drawn observations whose label also covers bar $t$ . Drawing an observation that overlaps what is already in the bag drives $\bar u_i$ down, so the next draw prefers something disjoint — this is what stops the standard IID bootstrap from silently resampling the same overlapping event $B$ times. Probabilities are recomputed after every draw. See sampling for the uniqueness machinery.

Usage Examples

Rust

Instantiate SB bagging classifier

use openquant::sb_bagging::SequentiallyBootstrappedBaggingClassifier;

// The single constructor argument is `random_state` — NOT the ensemble size.
// n_estimators defaults to 10 and has to be set explicitly.
let mut bag = SequentiallyBootstrappedBaggingClassifier::new(42);
bag.n_estimators = 100;
bag.oob_score = true;

println!("{} estimators, seed {}", bag.n_estimators, bag.random_state);

API Reference

Python API

sb_bagging.fit_predict_sb_classifier
sb_bagging.fit_predict_sb_regressor

Rust API

SequentiallyBootstrappedBaggingClassifier
SequentiallyBootstrappedBaggingRegressor
MaxSamples
MaxFeatures

Risk Notes and Caveats

Sequential bootstrap improves diversity under event overlap.
Tune max_samples/max_features with out-of-sample monitoring.