The Market Fraction Hypothesis under Different Genetic Programming Algorithms

In a previous work, inspired by observations made in many agent-based financial models, we formulated and presented the Market Fraction Hypothesis, which basically predicts a short duration for any dominant type of agents, but then a uniform distribution over all types in the long run. We then proposed a two-step approach, a rule-inference step, and a rule-clustering step, to test this hypothesis. We employed genetic programming as the rule inference engine, and applied self-organizing maps to cluster the inferred rules. We then ran tests for 10 international markets and provided a general examination of the plausibility of the hypothesis. However, because of the fact that the tests took place under a GP system, it could be argued that these results are dependent on the nature of the GP algorithm. This chapter thus serves as an extension to our previous work. We test the Market Fraction Hypothesis under two new different GP algorithms, in order to prove that the previous results are rigorous and are not sensitive to the choice of GP. We thus test again the hypothesis under the same 10 empirical datasets that were used in our previous experiments. Our work shows that certain parts of the hypothesis are indeed sensitive on the algorithm. Nevertheless, this sensitivity does not apply to all aspects of our tests. This therefore allows us to conclude that our previously derived results are rigorous and can thus be generalized.