Tim Maly of Quiet Babylon, has a thoughtful post about the way that the press deals with a financial market in which most trading is automated. Maly criticises the press for attributing unexpected market behaviour to human responses to news. He notes that 60% of trading is automated, suggesting therefore that software glitches are a far more reasonable explanation for such unexpected events. Consistent with some of his commenters, I disagree with this conclusion.
Just because machines make 60% of the trades doesn’t mean that they constitute 60% of the information coming into the system. In fact, as Maly rightly point out, battlebots (as they are called) do not do a natural language parsing of the news: they respond only to the state of the system itself. This means that the entire information input of the system is the human traders. The reason for these perturbations is the intersection of human traders who respond to news, and machines which respond to the human traders.
The machines may not be well-behaved. Their programmers have given them complex instructions, and the instructions of different machines may not be particularly different. It’s quite possible that some particular – apparently benign – stimulus might cause several machines to react in very similar ways. It’s possible that this group of initial actors may be very large subset of the machines. Moreover, other machines, and human traders which don’t respond directly to that stimulus, may respond similarly to the synchronised activity of the initial actors.
The market is a very complex system. It is very difficult to instruct a machine to act sensibly under all possible circumstances. The machine’s software will is designed under some particular model of the market’s behaviour. Whatever model is used, there is always the possibility of black swan events, which are unaccounted for under this model. Even with an incredibly sophisticated model, the machine’s software may not be a rational optimiser under all cases. Even if the behaviour of one machine may be somewhat predictable, the emergent behaviour of a large group of machines – like that of a flock of birds, or a school of fish – may be as difficult to predict as that of the market itself.
This synchronised emergent behaviour may result in poor collective action. Given that these events are quite possibly outside the model used to create the machines’ software, there may even be non-optimal individual action. If this is what Maly means by a “software glitch”, then we have a quibble of language, and perhaps causality, but little more than that. When I hear “software glitch”, I think of a rounding error, buffer overflow, or similar low-level problem, not a problem with the high-level logic of the model which the machine uses. After all, this sort of logic error is not really much different when acted on by a machine than by a human. Machines are faster, and less-likely to second-guess their decisions, but their decisions are likely to be much the same.
Whatever happens, some small set of human traders (constituting no more than 40% of trades) read the news, and take some action. We may not know exactly how their minds work; we may not even consider their actions to be particularly odd. Whatever their decisions are, they create some sort of mild change or perturbation in the market. A large group of machines faithfully and dutifully execute their well-programmed, but incompletely-reasoned instructions. This causes a significant market shift, amplified by more traders and machines acting rationally according to their models. None of this is caused by a rounding error, or a fat finger: just a series of actors duly going by their playbooks.
However, the initial impetus comes from the human traders. Their actions are the only ones influenced by news. That influence may be complex and tenuous. Human psychology may be as complex as trading software, if not more so. The simple explanations reported by the press may be speculative fiction designed to sell copy. Nonetheless, their core assumption is sound. People make trades in response to information. Only after that do the machines exhibit complex, unpredictable, emergent behaviour which – depending on the stimulus – may yield aberrant results.