Financial decision-making is a complex, even chaotic field of human endeavor that is increasingly being addressed by machine learning using agent-based models. The benefits of being able to use reliable predictive heuristics in financial decision-making are obvious, but the path to that goal is not. Past and current research has moved from social sciences and biology to economics, and multiple approaches are needed to identify basic theories that can be explored for greater development. This research proposal aims to evaluate past and current research using a systematic review and meta-analysis to identify promising approaches and define nascent basic theories in the field of machine based financial decision-making.
Keywords: machine learning, agent-based models, financial decision-making.
In 2008, a global financial meltdown rocked the world’s economy. Labeled as “The Great Recession,” the inevitable search for culprits and factors in the debacle revealed that the technology in use within financial markets was often poorly understood. Worse, the underlying principles of the use of said technology were also poorly defined. While technology was not a major factor in the causes leading up to The Great Recession (Fievet & Somette 2018; Caiani et al 2016; Shen & Tzeng 2015), it made for a neutral scapegoat and a field of opportunity to develop enhanced, more adaptable systems for dealing with the complexities of financial markets.
This proposal aims to define the state of the art in machine learning systems and agent based models used to create simulations for broader and deeper understanding of the financial markets in order to identify promising research avenues and trends that could have a significant if not dramatic impact on global finance in the near future.
Background and Rationale – 1,200
Machine learning is a sub-field of computer programming that intersects with artificial intelligence development (Jordan & Mitchell 2015, p. 257). The goal is to help computers and networks became more capable at analyzing non-objective factors in order to structure knowledge and self-learning (Shen & Tzeng 2015, p. 3-5). Computers and networks (systems combining multiple electronic devices) have already shown high proficiency levels in dealing with multiple attribute decision-making (weighing factors, as in medical diagnosis) and multiple objective decision-making (correlating treatment options based on the diagnosis), along with the capacity to move beyond initial programming to adopt new methods based on acquired information (Shen & Tzeng 2015; Burrell 2016).
However, the new frontier, and the area that machine learning has the potential to make the largest impact in, is multiple criteria decision-making, as this method applies to a wider range of fields dominated by “fuzzy logic” or “opacity” (Burrell 2016, p. 2). Financial markets are endeavors where humans invest enormous resources of time and certainly money, and where human intelligence alone is unable to entirely grasp the complexities involved, thus the promise of machine learning is to establish a path to competitive advantage or market domination.
Agent-based models are computer simulations used to emulate the actions and interactions of autonomous individual units or groups in order to evaluate their impact on the entire system (Smith & Conrey 2006, p. 6). Their initial use was in analyzing potential results of models in the social sciences and biology, using complex systems theory, evolutionary programming, and eventually game theory to apply to more direct interactions such as business negotiations and economics (Rai & Henry 2016, p. 2-3). The basic process is to define the system factors, conditions, drivers, and outputs (range of results) and introduce semi-random variabilities (within established parameters) to run thousands of iterations (van der Hoog 2016, p. 4-5). The evaluation is then made based on the median (most frequent outcomes) and the average of all results, adjusting the basic metrics to “zero” the model as close to possible as reality (Smith & Conrey 2006, p. 6).
The discipline of machine learning has two fundamental goals: “1) How can one construct computer systems that automatically improve through experience? and 2) What are the fundamental statistical-computational-information-theoretic laws that govern all learning systems, including computers, humans, and organizations?” (Jordan & Mitchell 2015, p. 255). In order to achieve those goals, researchers across fields as varied as linguistics, cancer, biochemistry, and economics have relied on agent-based models to create the “laboratory scenarios” in which the simulations interact, improving algorithms to bridge gaps between objective data and well-founded conclusions increasingly based on intangibles (Burrell 2016, p. 9). Thus the transition from attributes to objectives to criteria, in other words, from the facts to the intended results to the way in which the intended results can or will emerge (Shen & Tzeng 2015).
It is worth noting that what was once a programming issue (telling the computer or system the basic rules of its simulation) has given way to a more organic approach in which the system is told the initial state and the results and it “teaches itself” how the targeted system works (Reid 2017). The basis for this paradigm shift is that machine learning has proven a self-learning capability beginning with simply the barest of instruction sets (saving time and money), and that as no system is perfect, no programming instruction set could ever account for imperfections, therefore the best learning system is one that “learns what happens” to then attempt a structuring (Reid 2013).
As applied to financial decisions, the complexities of stock markets, foreign exchange transactions, and other aspects of national and global economics are an obvious field for machine learning, not only because of the almost-chaotic “feel” of the complexity, but because even a modest success in anticipating a result within that chaos could represent enormous benefits (Ghosh & Raju Chinthalapat 2014, p. 1-2).
Aside from the gigantic amounts of data produced and used within financial markets, a torrent difficult to parse and analyze efficiently, another major reason financial decision-making has yet to benefit greatly from widespread digital technology is that past performance bears (apparently) little relationship with current results (Neri 2011, p. 4). This creates scenarios in which a majority of conditions (combined basic and ongoing factors) form the basis of a pattern, but small internal and external factors can create outlier results that exceed present predictive capabilities (Hellenes 2016, p. 14).
However, this is not a situation that only appears in financial decision-making: it is part and parcel of systems where humans have frequent direct interventions and in natural systems, such as ecological niches, genetics, and malignant cell division (van der Hoog 2016, p. 2; Smith & Conrey 2006, p. 8-9; Wang et al, 2015). While agent-based models can narrow the gap of understanding variables within the system (Lamperti et al 2017), the goal of achieving deeper understanding will require establishing a baseline, a fundamental benchmark from which the basic theory can be tested and refined over time (Caiani et al 2016, p. 6-7).
Unlike machine learning that teaches itself to play a game, where the level of learning can be measured by a clear victory or defeat, machine learning that can enhance and strengthen predictive financial decision-making must prove itself by understanding the system enough to make predictions, and then contextualize the predictions within a range of certainty (or uncertainty) (Ghosh & Raju Chinthalapat 2014, p. 4). The development sequence is thus one where multiple approaches seek to identify and define how the system processes interact (attributes and objectives), but also identify and define the internal (system-based) and external (human-based intervention) criteria that have significant impact on the way in which the processes “play out” towards outcomes (Neri 2016).
If one were to diagram a model of the research, it would place the abundant data of financial decision-making on one side, with the largely-unseen rational and emotional factors that go into financial decisions on the other, with the agent-based model of machine learning as a bridge-builder between the two sectors (Hellenes 2016). New approaches are thus more focused on finding externalities that have been overlooked but that may play a role in financial decision-making, such as disclosures in the media (Chiong et al 2018). In addition, trends tend to build on themselves and create other effects, largely unexpected, and these emergent phenomena can cast incisive light on the complexity within financial decisions (Fiecet & Somette 2018).
With the development of more accurate “fuzzy maps,” combining knowledge with a broader capacity for machine learning to “think” in innovative ways (Mehryar et al 2018), the need is to define productive research paths that can expeditiously assist in developing reliable machine-based analysts and predictors in financial markets. As yet to neat-analytical study of such approaches has been made, marking this research study as a much-needed pioneer.
The proposed research questions are intended to frame and possibly define the current state of the art in machine learning using agent-based models for financial decision-making. The study will do this by identifying promising areas of research as indicated by the use of fields other than economics, the number and type of subsequent studies emerging from past research, and how the results of current research have helped close the predictive gap in financial decision-making.
The proposed research questions are:
What are the most common approaches in the research of machine learning using agent-based models for financial decision-making?
What fields of study outside of economics have been used to develop research of machine learning using agent-based models for financial decision-making?
What research approaches have yielded the greatest amount of follow-up or sequential studies of machine learning using agent-based models for financial decision-making?
The proposed research is a qualitative mixed-method design using a systematic review of research carried out in machine learning using agent-based models for financial decision-making, then using these results to produce a meta-analysis of the state-of-the-art in the field to determine promising avenues for further studies or new approaches that could yield significant results.
An extensive search through multiple databases will cover studies carried out between 1980 and 2018 (inclusive) using keywords such as: machine learning, agent-based models, financial decision-making, artificial intelligence, self-learning, chaos theory, game theory, and others. The goal is to cast the widest net possible so as to evaluate the broadest range of research approaches already completed (Cresweel & Creswell 2018).
Using pre-determined criteria, studies will be screened for relevance and validity, sing outside experts to validate the final selection. Once the basic analysis has been completed (number of studies, approaches, external fields of research, results, etc.) the meta-analysis will seek to pinpoint effective approaches, promising avenues for new or extended research, and the current level of decision-making capability based on machine learning.
Plan of Work
Once approval for the study is granted, the database search can begin immediately and is expected to take a week. Screening of the studies and evaluation the selections could take up to two weeks. The initial systematic review would take a week, with another week for the meat-analysis. Writing the final paper will take a week as there may be a need to do further research identified in the meta-analysis that was not fully explored (for example, an applicable study done in a field that was left out of the database search). From start to completion the research is expected to take 6-7 weeks.
EffectivePapers.com is a professional research proposal writing service committed to writing non-plagiarized custom proposals of top quality. All research proposals and thesis proposals are written from scratch by PhD academic experts you can hire online. Just proceed with your order, and we will find the best expert for you!
Burrell, Jenna. “How the machine ‘thinks’: Understanding opacity in machine learning algorithms.” Big Data and Society, 2016, Vol. 3, pp. 1-12. DOI: 10.1177/2053951715622512
Caiani, Alessandro et al. “Agent based-stock flow consistent macroeconomics: Towards a benchmark model.” Journal of Economic Dynamics and Control, 2016, Vol. 69, pp. 375-408. https://doi.org/10.1016/j.jedc.2016.06.001
Chiong, Raymond et al. “A sentiment analysis-based machine learning approach for financial market prediction via news disclosures.” Proceedings of the Genetic and Evolutionary Computation Conference, July 2018. DOI: 10.1145/3205651.3205682
Creswell, John & Creswell, J. David. Research Design: Qualitative, Quantitative, and Mixed Methods Approaches, (5e). Boston: SAGE Publications, 2018.
Fievet, Lucas & Somette, Didier. “Calibrating emergent phenomena in stock markets with agent based models.” PLoS One, 2018. https://doi.org/10.1371/journal.pone.0193290
Ghosh, Pradeep & Raju Chinthalapat, V.L. “Financial Time Series Forecasting using Agent Based Models in Equity and FX Markets.” 2014 6th Computer Science and Electronic Engineering Conference (CEEC). DOI: 10.1109/CEEC.2014.6958562
Hellenes, Herman. “A bridge between statistical learning and agent based modelling in stock market predictions.” Masters Thesis, 2016, Department of Management, Technology, and Economics, ETH Zürich.
Jordan, Mark & Mitchell, Tom. “Machine learning: Trends, perspectives, and prospects.” Science, 2015, Vol. 349, Issue 6245, pp. 255-260. DOI: 10.1126/science.aaa8415
Lamperti, Francesco et al. “Agent-Based Model Calibration using Machine Learning Surrogates. Économie et finance quantitative, 2017. HAL I.D.: hal-01499344
Mehryar, Sara et al. “Making use of Fuzzy Cognitive Maps in Agent-Based Modeling,” 2018. Department of Urban and Regional Planning and Geo-Information Management, Faculty of Geo-Information Science and Earth Observation, Conference Paper. Web. https://research.utwente.nl/en/publications/making-use-of-fuzzy-cognitive-maps-in-agent-based-modeling Accessed: 21 October 2018.
Neri, Filippo. “Learning and Predicting Financial Time Series by Combining Natural Computation and Agent Simulation.” EvoApplications 2011: Applications of Evolutionary Computation, pp 111-119. https://doi.org/10.1007/978-3-642-20520-0_12
Rai, Varun & Henry, Adam. “Agent-based modelling of consumer energy choices.” Nature Climate Change, 2016, pp. 1-7. DOI: 10.1038/NCLIMATE2967
Reid, Stuart. “Computational Decision Making Methods,” 2017. Web. http://www.turingfinance.com/computational-decision-making-for-finance/ Accessed: 21 October 2018.
Reid, Stuart. “Perfect Imperfection, Agent Based Models,” 2013. Web. http://www.turingfinance.com/perfect-imperfection-agent-based-models-abm/ Accessed: 21 October 2018.
Shen, Kao-Yi & Tzeng, Gwo-Hshiung. “An Emerging Trend in Multiple Rule-Based Decision Making by Rough Machine Learning.” 2015 International Conference on Fuzzy Theory and Its Applications, Yilan/Taiwan.
Smith, Eliot & Conrey, Frederica. “Agent-Based Modeling: A New Approach for Theory-Building in Social Psychology.” Personality and Social Psychology Review, 2007, Vol. 11, No. 1, pp. 87-104. https://doi.org/10.1177/1088868306294789
Van der Hoog, Sander. “Deep Learning in Agent-Based Models: A Prospectus,” 2016. Working Papers in Economics and Management, Universität Bielefeld. ISSN: 2196-2723.
Wang, Zhihui et al. “Simulating cancer growth with multiscale agent-based modeling.” Seminars in Cancer Biology, 2015, Vol. 30, pp. pp. 70-78. https://doi.org/10.1016/j.semcancer.2014.04.001