A carregar...
Publicação
Hyperparameters optimization on neural networks for bond trading
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 6.85 MB | Adobe PDF |
Autores
Orientador(es)
Resumo(s)
Artificial Neural Networks have been recently spotlighted as de facto tools used for classification. Their
ability to deal with complex decision boundaries makes them potentially suitable to work on trading
within financial markets, namely on Bonds. Such classifier faces high flexibility on its parameters
in parallel with great modularity of its techniques, arising thus the need to efficiently optimize its
hyperparameters. To determine the most effcient search method to optimize almost the majority of the
Neural Networks hyperparameters, we have compared the results obtained by the manual, evolutionary
(genetic algorithm) and random search methods. The search methods compete on several metrics from
which we aim to estimate the generalization capability, i.e. the capacity to correctly predict on unseen
data. We have found the manual method to present better generalization results than the remaining
automatic methods. Also, no benefit was found on the direction provided by the genetic search method
when compared to the purely random. Such results demonstrate the importance of human oversight
during the hyperparameters optimization and weight training phases, capable of analyzing in parallel
multiple metrics and data visualization techniques, a process critical to avoid suboptimal solutions
when navigating complex hyperspaces.
Descrição
Project Work presented as the partial requirement for obtaining a Master's degree in Statistics and Information Management, specialization in Risk Analysis and Management
Palavras-chave
Artificial Neural Networks Bonds Trabalhos de projecto