Deep Learning for Portfolio Allocation
In this work, we combined AI with an existing quantitative portfolio allocation model. In particular, we used the prediction of a Deep Neural Network as “investor views” in the Black-Litterman allocation model. This MSc thesis won the SIAT Technical Analyst Award 2019.
Abstract:
Financial markets are complex, non-linear, dynamic and adaptive systems that require advanced tools and techniques to capture the relationships that emerge within them (Brock and De Lima, 1995). For many years, finance relied on traditional statistical techniques and models based on assumptions often far removed from real-world market behaviour in order to extract insights from data.
Machine Learning, a subfield of Artificial Intelligence, promises to transform the understanding of financial markets by enabling researchers to apply modern techniques similar to those used in the hard sciences, capable of identifying highly dimensional and non-linear relationships.
When selecting the optimal allocation of resources across asset classes, investors have traditionally relied on newspaper articles, analysts’ reports and economic indicators, often depending heavily on personal intuition and judgement. Behavioural finance has recently highlighted how emotionally driven and biased evaluations have led to unrealistic portfolio allocations, with classical models from Markowitz (1952) to Black-Litterman (1992) often amplifying estimation errors.
Over recent decades, advances in computing power, the increasing size of datasets and new developments in data analysis have enabled investors to integrate investment strategies with more advanced tools, supporting more rational and data-driven decision-making.
Deep Learning, when applied to time-series forecasting problems, has achieved remarkable results and is now considered one of the most effective methods available for approximating complex functions (Hill, 1996). Its strong out-of-sample predictive capability enables the future behaviour of financial assets to be estimated more effectively than through human judgement or traditional regression methods.
It is expected that the asset management industry will be significantly reshaped over the coming decade (PwC, 2017). Artificial Intelligence represents an opportunity to revitalise a declining business model while improving operational efficiency and reducing costs.
The central aim of this thesis is to investigate the use and future potential of Artificial Intelligence in Asset Management and to implement an asset allocation system capable of outperforming traditional benchmarks by using signals generated by an Artificial Neural Network.
The asset allocation framework adopted during the experimental phase of the thesis is the Black-Litterman model, a cornerstone of modern portfolio theory. By incorporating forecasts generated by a neural network rather than human judgement or simple estimators, the model is able to deliver more stable and improved results. Furthermore, instead of relying on the Capital Asset Pricing Model to determine the equilibrium starting portfolio, the framework developed in this work employs a (Hierarchical) Risk Parity portfolio generated through a Machine Learning algorithm.