State of the Art Review for Applying Computational Intelligence and Machine Learning Techniques to Portfolio Optimisation

Computational techniques have shown much promise in the field of Finance, owing to their ability to extract sense out of dauntingly complex systems. This paper reviews the most promising of these techniques, from traditional computational intelligence methods to their machine learning siblings, with particular view to their application in optimising the management of a portfolio of financial instruments. The current state of the art is assessed, and prospective further work is assessed and recommended

💡 Research Summary

The paper provides a comprehensive review of the most promising computational intelligence (CI) and machine learning (ML) techniques applied to portfolio optimisation, positioning these methods within the broader context of modern finance. It begins by highlighting the critical role of portfolio allocation in both national and global economies and points out the limitations of traditional mathematical models, which often struggle to capture the non‑linear, high‑dimensional nature of financial markets. The authors argue that CI and ML, which rely on observation‑driven learning rather than explicit analytical formulations, offer a pathway to overcome these shortcomings, especially given the rapid growth in computational resources over the past two decades.
The background section underscores the relevance of portfolio optimisation, referencing the sub‑prime crisis as a stark illustration of the fragility of conventional risk models. It also notes that the sheer volume of financial data and the increasing complexity of investment instruments demand more sophisticated algorithmic solutions.
In the core technical review, three major CI methods are examined in depth:

1. Genetic Algorithms (GA) – The paper describes GA as a population‑based search that evolves candidate solutions through selection, crossover, and mutation guided by a fitness function. GA’s strength lies in handling multiple, possibly non‑convex constraints typical of multi‑asset portfolios. However, the authors caution that fitness design is highly subjective, and GA performance is sensitive to parameters such as mutation rate and population size, making practical deployment costly in terms of tuning.
2. Particle Swarm Optimisation (PSO) – PSO is presented as a velocity‑and‑position update scheme inspired by swarm behaviour. It converges quickly on continuous optimisation problems and has been successfully applied to weight allocation in portfolios. The paper notes the risk of premature convergence to local optima and the need for careful parameter selection (inertia weight, cognitive/social coefficients) to maintain exploration.
3. Artificial Neural Networks (ANN) – The review covers feed‑forward multilayer perceptrons, convolutional and recurrent architectures, emphasizing their universal function approximation capabilities. ANNs have been used for price forecasting and risk factor extraction, yet the authors highlight persistent challenges: over‑fitting, sensitivity to non‑stationary market regimes, and the difficulty of capturing extreme events due to the assumption of smooth, differentiable mappings.
   The machine‑learning segment focuses on Reinforcement Learning (RL), distinguishing on‑policy (e.g., SARSA) from off‑policy (e.g., Q‑learning, DQN) approaches. RL’s appeal is its ability to learn optimal trading or rebalancing policies directly from reward signals without an explicit model of market dynamics. Temporal‑Difference (TD) learning is discussed as a means to update value estimates online, enabling real‑time adaptation. Nevertheless, the paper stresses that RL suffers from stability issues in high‑dimensional continuous action spaces, and the design of reward functions that faithfully represent the risk‑adjusted return trade‑off remains an open problem.
   The authors then revisit Portfolio Theory. Modern Portfolio Theory (MPT) is summarised as the mean‑variance framework that defines the efficient frontier under the assumption of normally distributed returns and independent price movements. The paper critiques these assumptions, citing empirical evidence of heavy tails, autocorrelation, and market jumps that invalidate the Gaussian model. Post‑MPT extensions—such as using semi‑variance, Value‑at‑Risk (VaR), Conditional VaR, and other downside‑risk measures—are presented as attempts to better capture real‑world risk. The discussion on risk quantification introduces alternative metrics like fractal indices and Mandelbrot‑style volatility measures, arguing that traditional standard deviation underestimates true exposure.
   In the optimisation application section, the review surveys empirical studies that combine GA, PSO, ANN, and RL (often in hybrid configurations) to solve portfolio selection, asset‑allocation, and dynamic rebalancing problems. Results generally show improved risk‑adjusted returns compared with classic quadratic optimisation, but the authors note a lack of systematic evaluation regarding data preprocessing, hyper‑parameter optimisation, computational cost, and robustness to out‑of‑sample market conditions.
   The conclusion synthesises the findings and outlines four key research gaps: (1) robust modelling of non‑Gaussian, non‑stationary financial data; (2) principled reward design and policy generalisation in RL; (3) scalable, low‑latency computing infrastructures for real‑time deployment; and (4) extensive back‑testing and live‑trading validation to bridge the gap between academic prototypes and industry practice. The paper calls for the development of hybrid frameworks that integrate the strengths of CI, ML, and advanced risk metrics, along with open‑source toolchains that can be adopted by practitioners.