Detecting Speculative Bubbles Created in Experiments via Decoupling in Agent Based Models

Abstract

Proving the existence of speculative financial bubbles even a posteriori has proven exceedingly difficult[1-3] so anticipating a speculative bubble ex ante would at first seem an impossible task. Still as illustrated by the recent turmoil in financial markets initiated by the so called subprime crisis there is clearly an urgent need for new tools in our understanding and handling of financial speculative bubbles. In contrast to periods of fast growth, the nature of market dynamics profoundly changes during speculative bubbles where self contained strategies often leads to unconditional buying. A critical question is therefore whether such a signature can be quantified, and if so, used in the understanding of what are the sufficient and necessary conditions in the creation of a speculative bubble. Here we show a new technique, based on agent based simulations, gives a robust measure of detachment of trading choices created by feedback, and predicts the onset of speculative bubbles in experiments with human subjects. We use trading data obtained from experiments with humans as input to computer simulations of artificial agents that use adaptive strategies defined from game theory. As the agents try to maximize their profit using the market data created by humans, we observe certain moments of decoupling where the decision of an agent becomes independent of the next outcome of the human experiment, leading to pockets of deterministic price actions of the agents. Decoupling in the agent based simulations in turn allows us to correctly predict at what time tb the subjects in the laboratory experiments have entered a bubble state. Finally in one case where the subjects do not enter a permanent bubble state, our method allow us at certain special moments to predict with a 87% success rate an unit move of the market two time steps ahead.