Stacking Neural Network Ensembles for Market Direction Classification

Summary

This article studies stacking as a way to combine predictions from an ensemble of extreme learning machine classifiers. The base models produce outputs on separate training and evaluation samples; those outputs become features for a second-level combiner. The experiment compares combiners built from fully connected neural networks and examines different input sets, including a pruned ensemble and the full ensemble.

The workflow covers quote preparation, outlier handling, normalization, predictor selection, fitting base models, and generating the metamodel's inputs. Reported comparisons suggest that simple averaging or majority voting is fast and performs reasonably, while a small neural network using a softmax combination reduces bias without a noticeable variance increase. More complex combiners did not improve the results, and a Bayesian variable-selection logistic regression also performed well. The article acknowledges residual noisy labels and suggests sequence-aware models as future work. Results are tied to the described data and setup, with no evidence here of live trading performance or broad out-of-sample robustness.

Key ideas

• Stacking trains a second-level model on predictions produced by base classifiers.
• The quality of the metamodel depends on how its training inputs are constructed from ensemble outputs.
• The experiments compare pruned and full ensemble outputs as combiner inputs.
• Simple averaging and voting are computationally light baselines, while a neural combiner can reduce bias.
• The reported classification results leave unresolved noise and do not establish live trading profitability.

Tags

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