HFT Myths

Quote from rwk:

achieving alpha starts with intuition, and that takes experience.
More...
I'm going to use this line when they try to give this old geezer's job to hotshot college grads.
 
Quote from hft:

My definition of backtesting might not match yours. I define it strictly as simulating how a strategy performs through historical data. It's really hard to do this well with a high-frequency market-making strategy. Instead of that, I rely other forms of analysis (and admittedly 'educated guesses').

- Run blind: Simply rest small size and see if it makes money
- Trial and error: Tweak some settings and run it concurrently with a control strategy to see if it's more effective.
- Post-trade analytics: Analyze the trades you got into in order to improve the strategy.
- Historical data analytics: Analyze historical market data, but not necessarily simulated fills as in a typical backtest.

Some of that is quite vague, which is both to not reveal secrets as well as indicative of the ad-hoc nature of the process anyway.
More...


Thanks, HFT. It seems that our definitions of backtesting are identical. I presume that the hardness of strict backtesting is the lack of trustable market impact model. Is that correct, or are there other difficulties?
What strikes me is that with all these alternative ideas (and without strict backtesting) it should be hard to effectively optimize order placement strategies on past data.

You said that that only a minority of the MM players use strict backtesting. Are these the big and more sophisticated players that do manage to create an impact model? Or are these the smaller players who try to play without impact?
 
Quote from lastlook:

Thanks, HFT. It seems that our definitions of backtesting are identical. I presume that the hardness of strict backtesting is the lack of trustable market impact model. Is that correct, or are there other difficulties?
What strikes me is that with all these alternative ideas (and without strict backtesting) it should be hard to effectively optimize order placement strategies on past data.

You said that that only a minority of the MM players use strict backtesting. Are these the big and more sophisticated players that do manage to create an impact model? Or are these the smaller players who try to play without impact?
More...

Not sure if size/competency have as much to do with it as much as belief in and approach to the process. The strict MM'ing backtesters that I've personally known have all failed. The ones that I've heard of being successful at it presumably have good models for not just market impact, but many other variables like latency-distribution, queue position adjustments, exchange-specific detailing, etc. Some just don't think market impact alters the backtest result enough to write the backtest results off completely. Others believe in the effectiveness of their market impact models. I think more important than market impact is having good latency distribution, queue position, and exchange-specific models. Some fairly complex and successful firms/traders have tried to do all of these things, myself included. In the end I've always concluded that the variability that they introduce skew the backtest results enough to make them ineffective metrics of profitability and strategy effectiveness. If you understand the limitations you can gain useful insight from backtesting, but it can also lead to some very very bad conclusions being drawn.

Someone mentioned earlier that backtesting can at least rule out bad ideas, which I agree with. You can tune backtest settings to be very optimistic and if that doesn't make money you can dump potential strategies very quickly.
 
Quote from elitetradesman:

Thanks. I suppose you use marketable limit orders for that?
More...
Yep. However, 'marketable' is a very fickle term in HFT. But yeah, basically there is some price you're willing to buy/sell for and that's what you tag the order with.
 
What do you synchronize local time with? Exchange time, brokerage time, or UTC?

Do you synchronize local time at the OS level or the trading platform level?

Thanks again.
 
You must log in or register to reply here.
Back
Top