Exploring the "robot phase transition'' in experimental human-algorithmic markets

Johnson et al. (2012) recently argued that analysis of millisecond-by-millisecond stock-price movements between 2006 and 2011 suggests the existence of a step-change or phase transition in the dynamics and behaviour of financial markets, in which human traders and automated algorithmic 'robot' trading systems freely interact. Above a particular time-threshold, humans and algorithmic systems trade with one another; below the threshold, there is a sudden change to a market in which humans cannot participate and where all transactions are robot-to-robot. We refer to this abrupt system transition from a mixed human-robot phase to an all-robot phase as the 'robot phase transition'. At sub-second timescales, below the robot transition, Johnson et al. argue that the robot-only market exhibits 'fractures' - ultra-fast swings in price akin to mini flash-crashes - that are undesirable, little understood, and intriguingly appear to be linked to longer-term instability of the market as a whole.

Here, we report on using a complementary approach to the historical data analysis employed by Johnson et al.: in March 2012 we conducted laboratory-style experiments where human traders interacted with algorithmic trading agents (i.e., robots) in minimal experimental models of electronic financial markets using De Luca's (2011) OpEx artificial financial exchange. Our aim was to see if correlates of the two regimes suggested by Johnson et al. occur in such laboratory conditions. Our results thus far do indeed indicate that when trading robots act on a super-human timescale, the market starts to fragment, with statistically lower human-robot interactions than we would expect from a fully mixed market. In contrast, when robotic trader agents are slowed to a thinking-and-reaction speed similar to that of humans, less fragmentation is observed. We tentatively conclude that this is evidence for the robot transition occurring in controlled experimental financial market systems.

The work reported here also explore the effects of increasing the degree of realism in the laboratory experiments: we find that some statistically significant effects may be consequences of constraints introduced to make the analysis of the experiment results easier, and we report on our discovery of a problem with earlier OpEx experiments that casts doubt on their results.