Abstract
Many investors and financial managers view portfolio optimisation as a critical step in the management and selection processes. This is due to the fact that a portfolio fundamentally comprises a collection of uncertain securities, such as equities. For this reason, having a solid understanding of the elements responsible for these uncertainties is absolutely necessary. Investors will always look for a portfolio that can handle the required amount of risk while still producing the desired level of expected returns. This article uses feature-based models to investigate the primary elements that contribute to the optimal composition of a specific portfolio. These models make use of physical analyses, such as the Fourier transform, wavelet transforms and the Fourier-Mellin transform. Motivated by their use in medical analysis and detection, the purpose of this research was to analyse the efficacy of these methods in establishing the primary factors that go into optimising a particular portfolio. These geometric features are input into artificial neural networks, including convolutional and recurrent networks. These are then compared with other algorithms, such as vector autoregression, in portfolio optimisation tests. By testing these models on real-world data obtained from the US stock market, we were able to obtain preliminary findings on their utility.