Bayesian Optimization of Deep Neural Network Trading Classifiers

Summary

This article describes a workflow for tuning a two-hidden-layer deep neural network used to classify market data. It prepares pretraining, training, validation, and test sets; removes predictors judged statistically insignificant in earlier work; and uses mean class F1 as the scalar objective for Bayesian optimization. The search covers network size, activation functions, dropout, and learning settings across pretraining and fine-tuning configurations. The model uses pretraining followed by backpropagation or resilient propagation, and the article outlines testing and forward-testing with selected parameters.

The author reports a 7–10% improvement in classification quality from hyperparameter optimization and recommends multiple optimization runs, with results used to seed later searches. The discussion also notes computational cost and proposes additional parameters and neural-network ensembles for future work. Results depend on the stated data preparation, split choices, and model setup; the excerpt does not establish that the classification gains translate into trading returns or persist across other markets and periods.

Key ideas

• Bayesian optimization selects DNN hyperparameters by maximizing mean F1 across classes.
• The workflow separates pretraining, fine-tuning, validation, testing, and forward testing.
• The search includes hidden-layer sizes, activation choices, dropout, and learning parameters.
• The author reports improved classification quality but does not demonstrate improved trading performance.
• Repeated optimization runs and staged searches are suggested, at the cost of additional computation.

Tags

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