Robo-advisor using genetic algorithm and BERT sentiments from tweets for hybrid portfolio optimisation

•We proposed two robo-advisors capturing market conditions through Twitter sentiments.•Google’s BERT model was used to convert tweets to sentiments.•Genetic algorithm was used to optimise the models for different objectives.•The proposed models achieved superior portfolio performance. Robo-advisors...

Full description

Saved in:

Bibliographic Details
Published in:	Expert systems with applications Vol. 179; p. 115060
Main Authors:	Leow, Edmund Kwong Wei, Nguyen, Binh P., Chua, Matthew Chin Heng
Format:	Journal Article
Language:	English
Published:	New York Elsevier Ltd 01-10-2021 Elsevier BV
Subjects:	BERT Genetic algorithm Genetic algorithms Machine learning Optimization Portfolio optimisation Portfolio performance Social media sentiment analysis Twitter Weather Social media sentiment analysis Twitter BERT Genetic algorithm Portfolio optimisation
Online Access:	Get full text
Tags:	Add Tag No Tags, Be the first to tag this record!

Description
Summary:	•We proposed two robo-advisors capturing market conditions through Twitter sentiments.•Google’s BERT model was used to convert tweets to sentiments.•Genetic algorithm was used to optimise the models for different objectives.•The proposed models achieved superior portfolio performance. Robo-advisors are increasingly popular, with machine learning algorithms taking centre stage for researchers. However, classical financial theories and techniques, such as Constant Rebalancing (CRB) and Modern Portfolio Theory (MPT), can still be relevant by combining them with social media sentiments. In this study, we propose two novel models, namely Sentimental All-Weather (SAW) and Sentimental MPT (SMPT), which capture the up-to-date market conditions through Twitter sentiments via Google’s Bidirectional Transformer (BERT) model. Genetic Algorithm was used to optimise the models for different objectives including maximising cumulative returns and minimising volatility. Trained on tweets and the United States stock data from August 2018 to end December 2019, and tested on an out-of-sample period from January 2020 to April 2020, our proposed models achieved superior performance in terms of common measures of portfolio performance including Sharpe ratio, cumulative returns, and value-at-risk, compared to the following benchmarks: buy-and-hold SPY index, MPT model, and CRB model for an All-Weather Portfolio.
ISSN:	0957-4174 1873-6793
DOI:	10.1016/j.eswa.2021.115060