HFT trading is a fascinating subject. Nothing brings out creativity as when objective function is your real account balance.
There are two major things here, one is market making/ providing liquidity, where the HFT guys not necessarily make money from actual directional bets, but are paid by other clients to disguise large trades, acting as "buffers", since knowledge that a large trade is being done, is by itself very valuable, and small traders will front-run the price in front of a large entity establishing the position..
The other thing, is actually developing algorithms to trade thinly traded stocks and options, and doing price discovery... That is what killed Facebook IPO, 100'th of bots, issuing thousand of "probing" bid / ask changes per second, in order to see where that thing is going, which overwhelmed nasdaq completly.
For example you can observe it real time, in after hours session, on thinly traded stocks, how two algorithms programmed to remain top bid/ask by 0.01c, will walk the price down, chasing each other until pre-set limit. Or using a binary search you can discover algorithm's current limits, then walk the price up to that limit, and finally slam them by taking the trade, often saving up to a dollar in bid/ask spread, by outsmarting the bot.
Its an optimization problem, as every time algorithm uses a "probing" 100 share order, to see if someone will take it or not, they run the risk of someone taking the offer. So it becomes "is risk of going short 100 at price X, worth the extracted knowledge that there are agents willing to buy at price X"
Since these 100 share games are so frequent, other algorithms are getting smart and start react to only 200 shares, or to price*shares trigger.. That ups the stake, as the ultimate question becomes "how much is information worth"
Then add a level of indirection. If you can figure out what information algorithm is trying to extract.. The competing algorithm may actually hit the offers, not to make money, but to fool the algorithm doing the price probing, into thinking there are agents willing to buy at certain price.
Then you add options on top of that, and the fact that option time value before/after earnings, acts as closed energy system, with hidden information (the earnings), so whole fat tailed range of possible earnings, including complete blowout, or complete disaster is reflected in option time value, but once hidden information becomes public (earnings released), options time value suddenly drops, since some possibilities are eliminated.. Kind of reminds me of the probability function collapse thing.
Its a very fasinating field, only problem is anyone having any kind of success in it, is sitting tight and not publishing anything, or in fact may publish disinformation as to lead other researches on the wrong track.. So above may be completely bonkers and in fact written to mislead :-)
