1) I've really done a bit of everything, from low-level market connectivity development to high-level alpha research, management, and everything in between. 2) You can sum up the changes in HFT over the past decade by saying everything and everyone has gotten faster and smarter. Margins are much smaller as as a consequence. There are also many more players - the same piece of pie is being split up many more ways. Many exchanges that were more closed-off years ago (Asia, South America) have all catered to HFT firms now. Markets globally have become more intertwined. Many of these points are secondary to the major underlying drivers of change - the increase in speed and decrease in profit margins. 3) It varies too much by strategy type to really put a number on. A lot of time is spent improving existing strategies as well. In a group you might have 2/3 of the personnel working on existing strategies while 1/3 focus on new strategies. You can roll out a simple trade in a matter of hours, say to capitalize on large global events that create opportunistic trading conditions. Others can take months to research and implement. 4) Again, varies a lot by strategy. Generally speaking, as you increase complexity and decrease latency sensitivity, you increase the half-life of a strategy, at the expense of increased risk (lower sharpe) and longer development time.
Very interesting discussion. 1. Who are the big players in the HFT space outside of Getco, Virtu, HRT, Citadel, Tower? 2. On the equities side, what percentage of the trades are scratch, winners, and losers (either before rebate or after) and average profitability per share? 2. What percentage of your resting orders get filled? 3. What is your ratio of remove to add? 4. How do you decide when to remove? 5. How many shares/contracts do you trade on a typical day? 6. How do you size your orders? 7. What type of hardware are you using? FPGA? 8. Do you use third party software?
Glad to see a post from a fellow HFT'er . Good questions to ask, that's why I can't answer many of them . 1. There's quite a few that come to the top of my head. In no particular order: Jump, Teza, Spot, Peak 6, Optiver, IMC, Jump, Allston, Volant, XR, Ronin, EWT, Eagle 7, Wolverine, and the list goes on for a while longer. 2/2/3/4/5/6/7. Can't tell you. 8. No. We refuse to even pay for Windows or Office. It's open-source or in-house here.
How do HFT's solve the partial fill problem on the FX leg when trading ADRs or other dual listed assets across markets?
LOL - The guy who built the firm and correspondingly its culture must be a Steve Jobs brother hating MSFT from the core!!!! I respect you that you have volunteered to answer all the questions but I wonder why you are apprehensive and don't want to divulge 2/3/4/5/6/7? We were having discussion about security of IP of HFT when you co-locate in another thread. And the point was made by another HFT practitioner that its all about technology, complexity and details. http://www.elitetrader.com/vb/showthread.php?s=&postid=3792315#post3792315 So, I don't understand what makes you apprehensive about answering these questions? After all, you are not sharing any ip. Without knowing your firm name, these are just some numbers like fill ratios etc. Big Deal?! EDIT: Actually you have already answered 2 in post #5 on this thread. So, you have infact shared this information before since you considered it to be unimportant. Not trying to be a pr**k, but I love to point logical inconsistencies. Hope you are not getting offended Cheers.
1) In case, you agree that success in this business is all about extreme technology, millions to put into the technology and access to brilliant employees, would you be open to describe in detail couple of your strategies (either working currently or which have stopped working)? 2) Please see this academic paper on HFT - Paper developes a money losing strategy! http://www.ntuzov.com/Nik_Site/Niks_files/Research/papers/stat_arb/Ahmed_2009.pdf What is the resemblance between practical HFT strategies and HFT strategy like that described in this paper in terms of concepts and complexity? Again, Thank You Very Much!
Fair enough, I'll give numbers from one of my strategies from yesterday: 2% market share Remove vs. Add: ~ 1:1. Slightly more adds than removes due to fills. Duh. Removing modify's from the equation to keep things simple. % orders filled: ~10% Decide when to remove based on alphas Order Sizing: <= 10% of average top of book quantity Profit: ~ 1/10 of a tick net of all fees Hardware: Can't say anything more here. The reason not to reveal more is that it gives other HFT'ers a baseline to compare against. If someone was netting less than avg profit and trading more than my mkt share, they might try to back off a bit to increase the total PNL by increasing profit margin. Then again, I could be totally bluffing on all these numbers to throw people off-track. You're right, I mentioned avg % of winners/losers and PNL earlier, just to give you a rough idea of what's in the right order of magnitude (so that people don't assume that HFT"s make money on 90% of trades as I've seen written here and there).
I'm not quite smart enough to understand that full paper. But from what I can gather, they're predicting mid-price market movement. Conceptually it is identical to what we do. And in fact, when you're making markets you can use that model as-is to probably make some money since you're going to capture a bit of edge from capturing bid/ask. But you have to be fast enough to avoid getting picked off (i.e. reacting to your signal more quickly than other market participants). If I could give you a signal that predicted with 99% accuracy that the market will uptick (next mid-price will be 1 tick higher than the current), could you make money with it? Turns out I can. If you look at all instances where the ratio of quantity on the bid vs volume on the offer is greater than X, you can increase X until you get a probability approaching 100% of an uptick. But the catch is that a single uptick doesn't make you money. All that guarantees is that if you lift the offer, then you're very likely to be able to hit the bid at the same price later, which is a loser after fees. On top of that, as you increase X, you're going to become more latency-sensitive, vying for more competitive prices and your fill simulation will be thrown off, in addition to being subject to adverse selection of prices that aren't truly as attractive as your simple model predicts. As for a strategy in detail, here's one: Rest orders in symbol A at endpoint B for C ticks wider than the price that the same symbol is quoted on at endpoint D is quoting. When filled, hit A@D for a C tick winner. Easy as that. Feel free to change the variables to match anything traded in the world. A=EURUSD, B=Currenex, C=1, and D=Hotspot has worked for me in the past. And be sure to let me know how it goes!
Nothing out of the ordinary. Linux, (Bash/sed/etc), Python, and R for analysis. C++ for production. LibreOffice and Thunderbird for non-productivity.