In this paper we propose a new neural networks based regularization method requiring only 4 additional hyperparameter and that can be easily injected in any machine learning architecture. It is based on the use of auxiliary loss functions designed to appropriately learn data momenta. Our approach can be used both for classification and regression problems. A comparative analysis with real time series will be provided concerning cryptocurrency data, showing improvements in accuracy of about 5% with respect to existing approaches, without requiring additional training data or further parameters. The presented approach constitutes an innovative, new step towards the statistical moments oriented regularization scheme for statistical forecasting.
|Titolo:||Multitask Machine Learning for Financial Forecasting|
DI PERSIO, Luca (Corresponding)
|Data di pubblicazione:||2018|
|Appare nelle tipologie:||01.01 Articolo in Rivista|
File in questo prodotto:
|Multitask machine learning for financial forecasting.pdf||Versione dell'editore||Open Access Visualizza/Apri|