There is a great comment from a higher-frequency trader that was left on a post on Marginal Revolution. It’s a few comments down on that page. Here’s the background that he gives:
I work as a quant at one of the major high frequency trading firms, this paper is definitely one of the better academic works I’ve seen on the subject. I’ll add a little more. Generally the way that HFT works is by looking for a set of predictive signals in the market. Those signals are combined with liquidity and execution constraints to try to find the most profitable set of parameters after transactions costs are taken into account.
90% of these signals are fall into two major categories: 1) Looking at the price movements of related securities. A good example is SP500 versus Nasdaq. The correlation between the two is around 85% on a daily horizon, but over a horizon of 10 secs or so correlation is virtually zero. So when one moves a certain you bet that the other one will either follow or the first mover will fall back. 2) The other one is by looking at the state of the limit order book and it’s evolution through time. As a very simple example if say you have 20,000 size quantity on the bid and it’s been monotonically increasing and 5,000 size quantity on the ask and it’s monotonically decreasing then it’s very likely that the level on the ask will get wiped out first and the price will go up.
In general what these two add up to is trying to distinguish noisy trades versus signal trades. Speculators/investors/hedgers/etc. are the primary players in the market. Some of those trades contain high information (e.g. maybe a person with access to insider information buying up stock before some announcement), some of them contain virtually no information and are pure noise (e.g. granny liquidating some of her portfolio for monthly expenses). In a naive market with no HFT signals we have no way of assessing the informational content of individual trades, we only have an estimated aggregate or average informational content of trade. Market makers will set their spread and sizes according to this aggregated informational content.
But over any sample the estimated average informational content of trades will not be the same as the realized, for example one week might more than usual insider trading, one month it might make up a small fraction. There’s also a ton of path dependency when you work out the math, that amounts to pure randomness. Because of this securities will not perfectly track their “true price.” The deviation is still stationary, because the more out of line the prices get with the fundamentals the more speculators will step in and push it back. No one is smarter than the market 100% of the time so every time a fundamental speculator sees a price that’s too low/high there’s some chance that the market is right and his valuation is missing something and some chance he’s right. Speculators that aren’t very good are probably only going to be “beating the market” when the valuation on securities looks insanely out of whack or by distributing his portfolio over a wide range of perceived mis-valuations to reduce his volatility. Only the very best speculators are going to be able to get their fundamental valuations consistently right within a small margin of error. So without HFT/Stat Arb./technical trading/whatever you want to call it/etc. the thing that keeps securities from randomly drifting too far are fundamental speculators.
Basically what HFT is doing, instead of fundamentally valuing securites, determining the informational content of individual trades or small time frames, using the signals I mentioned earlier. A segment of the price evolution with high information content tend to look very different from noisy trades on the small scale, but when aggregated up lose this distinguishability. It’s almost symmetrical when you think about it. Fundamental speculators estimate a price for the security and trust in the reliability of the price evolution process in brining the market price to their estimated “true price”. HFT trusts in the reliability of the initial price as being the best estimate of the value of the security and tries to identify errors and miscalculations in the price evolution process.
There is a lot more to that comment there, and plenty of other worthwhile comments as well. Go give them a read!