Inspiration from evolutionary approaches: algorithms and challenges in finance

Advanced analytical techniques borrowed from a heterogeneous literature to extract valuable insights from data are gaining momentum within the finance community. This article briefly introduces how financial ecosystems are increasingly responsive to the application of biological and evolutionary algorithms aimed to analyze the behavior and dynamics of their participants.

Andrea Flori, Assistant Professor in Management and Finance at School of Management, Politecnico di Milano


In a popular 1973 work, Burton Malkiel showed how a blindfolded monkey throwing darts at the financial pages of a newspaper could select a portfolio which would perform just as well as one carefully constructed by experts (“A Random Walk Down Wall Street”, 1973), helping to fuel the debate on the possibility of extracting information from financial data and from the behaviour of market operators.

Financial markets are an arena where predictive techniques and algorithms attempt to challenge market efficiency using pattern identification.

Within this framework, numerous methodologies borrowed from various scientific fields have spread and combined to provide a new perspective for studying the dynamics of complex financial systems and the interdependencies that govern their participants’ relationships.

Market infrastructure, the role of information and the behaviour of the operators represent, therefore, some typical pillars employed to define such approaches of analysis within financial contexts.

In particular, approaches inspired by biological perspectives have captured the attention of many financial operators interested in a new generation of intelligent techniques of analysis and calculation that mimic human actions.

Hence, genetic algorithms and neural networks have pervaded financial literature and contributed to the dissemination of methodological innovations inspired by the biological evolution and human functioning. By relying on a multidisciplinary and computationally evolved perspective of analysis, these approaches have shown to obtain promising results compared to traditional methods of statistical analysis.

Specifically, genetic algorithms employ natural selection and genetics tools and perspectives to identify the best solution to problems. By mimicking biological evolution, an initial population is iteratively mutated and recombined to determine subsequent generations, so that changes with a desirable impact are retained in the gene pool of future generations in an attempt to converge towards optimal solutions. Each individual, i.e. candidate solution, is assigned a fitness value, and individuals with more promising characteristics are assigned a higher probability of mating to generate new individuals, i.e. potentially better performing solutions to problems, thus in line with the Darwinian theory of the “survival of the fittest”. Neural networks, instead, learn from data structure and process a signal through interconnected artificial neurons creating a configuration similar to the human nervous system. Each connection is associated with a weight that contains information about the input signal that inhibits or stimulates the signal that is communicated to the neurons to accomplish a specific task. External information is used as input and processed internally in one or more layers of analysis by activating specific neurons that transmit the signal to others before determining an output with a predictive accuracy that can be increased by a learning process of previous actions. Neural networks are a complex adaptive system which can change and adapt its internal structure based on the situation and information that crosses it.

These approaches are therefore flexible and able to adapt to new circumstances, possibly learning from past experience and reacting to stimuli coming from new signals in the system.

It is not surprising, therefore, that such techniques, either separately or combined with each other, are increasingly being applied in many financial areas, such as predictive market analysis and portfolio allocation rules, hedging of financial instruments, and robo-advisor applications.

With the recent increase in computational power and resources and their wide availability, advanced techniques for massive data analysis are, in fact, gaining momentum within the finance community, contributing to a rapidly growing body of literature that exploits, in addition to the techniques mentioned above, a large-scale use of statistical and deep learning concepts to identify patterns in financial markets, study the complex non-linear relationships between and within financial time series, and identify market anomalies.

In addition, repeated episodes of financial crisis with their far-reaching externalities and cascading effects on financial markets and the real economy have driven the study and provision of new tools to monitor and predict the spread of instability in financial systems and manage critical issues that may emerge.

However, adopting statistical learning approaches and deep learning techniques in the study of financial systems requires new paradigms, knowledge and practical skills necessary to develop a solid base of models and algorithms that are properly applied in the reference domains while exploiting the potential arising from a transdisciplinary approach to scientific investigation.

In the world of finance, such tools of analysis are of the utmost importance for future technological development and play a fundamental role in many financial ecosystems. Massive data analysis using advanced statistical and deep learning techniques is in great demand in many areas and in a wide range of applications, which include in addition to forecasting market trends, also for example the study of dependencies between financial systems, the approval of credit lines, the efficient management of financial resources, the detection of anomalies and frauds, and risk assessment.

These challenges seem to indicate therefore a new research perspective, at the crossroads between data mining, predictive analysis and causal modelling, which can allow us to exploit the strength of these algorithms of analysis and calculation to investigate economic and social problems of real and changing environments.