Tactical asset allocation

: a novel approach via machine learning returns prediction

Abstract

The thesis tries to develop a novel quantitative approach for the development of trading strate gies managed through Tactical Asset Allocation investment style. The work employs a returns prediction performed via supervised learning algorithms and studies whether they can help in the deliverance of superior returns compared with the classical buy-and-hold strategies. The models manage to predict the movement of the markets with an acceptable precision level and try to give hints for bet sizing, playing a key role in the weight's definitions. While most of the previously developed Tactical Asset Allocation strategies were performed through a historical analysis. The novelty of the approach comes from the incorporation of these forward-looking predicted values.In thisstudy, Multi-Layer Perceptron Neural Networks, Support Vector Machines, and Random Forests were used to predict the returns of two main market indexes, respectively S&P500 and Eurostoxx600. The models are used in their regressor form to get a continuous output, expected to be the true value of the returns for the following day. Based on this prediction, several trading strategies have been developed and tested.Results indicate that the proposed approach can give positive signals for what concerns return achievement and reward-to-risk ratio improvement. Nevertheless, due to the high dynamicity of the strategy, as implied by Tactical Asset Allocation hypothesis, transaction costs play a key role in final returns deliverance. All the trading strategies are performed considering the different outcomes of the models

Date of Award	2 Feb 2023
Original language	English
Awarding Institution	Universidade Católica Portuguesa
Supervisor	Dan Tran (Supervisor)