Bagging Neural Network Classifiers for Market Prediction

Summary

This article introduces classifier ensembles and explains bagging, or bootstrap aggregation, as a parallel method for reducing prediction variance. Each base classifier is trained on a bootstrap sample, and their outputs are combined through averaging or majority voting. The discussion places bagging among broader ensemble approaches, distinguishes parallel from sequential methods, and considers design choices such as diversity, ensemble size, classifier selection, and whether to use a fixed or trainable combiner.

For its implementation, the article uses extreme learning machines as fast, relatively simple base classifiers, varying samples and predictors to encourage diversity. It discusses pruning ensemble members and tuning their hyperparameters. The reported results indicate that optimization improves classification quality, averaging and voting perform similarly, and the ensemble’s predictive quality persists farther into the tested sequence than that of the compared deep network. Threshold calibration and stacking are proposed as further improvements. These findings depend on the article’s particular data, labels, and evaluation procedure, so they do not demonstrate profitability or general performance across markets.

Key ideas

• Bagging trains base classifiers on separate bootstrap samples and combines their outputs by averaging or voting.
• Low correlation and diversity among base models are important to the rationale for parallel ensembles.
• The article uses fast extreme learning machines and varies samples and predictors to produce different ensemble members.
• Hyperparameter tuning, threshold calibration, pruning, and stacking are discussed as ways to improve classification.
• The reported classification results are specific to the study and do not establish trading profitability.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.