Yes when I say on BATS I am "more likely" to get the lucky fill (when price just touches, not crosses, my price), it is still rare ... and at this level just posting an order, especially inside or near the NBBO, will affect the stock. But regardless if you are adding or taking liquidity ... you can reasonably simulate 100 share orders (leg in and out if you need to), but except for the most liquid stocks, it is almost impossible to simulate bigger orders (unless you rebuild the entire order book).
I'd posit the following. 1) Most hedge funds with very few exceptions do a good portion of their trading over the phone. When you are moving billions of dollars, latency really isn't the issue. 2) I don't think anyone at 5 ms would consider themselves low latency. There are several cheap off the shelf solutions that can beat that. Low I would say starts at 50-70 micros. Ultra low is below 15. Not many people are sub 5. And, as other people point out it's not like there are just fewer opportunities at slower speeds. They are completely gone. Like, all pure arb disappears by around 15 micros. Quick stat arb is pretty much gone by 70. After that you are pretty much in the same boat as everyone else. Your latency doesn't really matter and mostly likely won't improve your performance at all.
I'm curious as to how many of those contributing to this thread believe speed is a crucial part of their edge. I'm strictly a low tech directional trader (ES & CL) and would like to know if, for most in this thread, this discussion is more theoretical or more practicle.
Speed helps many strategies. I would guess your strategy would see no difference between 1ms and 200ms. It's all about trying to beat other guys looking at the same thing.
Speed helps me because about 40% of what I do is pure arb. Just can't really be done if you are over 15 micros. But for my 'slower' strategies - I run on a different system that about 100 micros. Could probably run at 200 milliseconds (2000x slower) with little discernable difference.
I think I remember from some of your earlier posts that you used to trade through IB. What type of latency do you think can be achieved there (trading equities) and how would you go about setting it up? Ballpark cost? I'm asking because I believe I have some strategies that are dead at 200ms, possibly worthwhile at 50ms, and quite good <=10ms. I currently run several automated strats through IB but with my current retail data feed they survive latency that is often measured in seconds. I gave Lime some serious consideration but ultimately decided that it wasn't worth moving everything over (at least not yet). If possible, I'd like to run some poor man's HFT through IB and see how that goes.
I can see that in the "pure arb" end of your business real speed is essential. Thanks. Makes the comments more concrete for me.
At least for smart routing it seems that >500 ms is not unusual for iB to send order acknowledgment ... (without your own latency). Good luck!
IB was a long time ago. So what I say might no longer be accurate. We used their FIX api and hosted a server in New Jersey which at the time was the fastest route to IB execution servers. Used NxCore for the data. I'm not sure what the delay is on NxCore data anymore. I know it's time-stamped to 25ms. Not sure where you are even getting 10ms data to backtest with. I have some memory of something crazy with NxCore like the data was routed to Utah then compressed and sent back to you. Not sure if that's the case anymore. Something to investigate. If that was the case from Jersey to Utah and back at a decent clip back in the day would be on the order of 75ms delay. So, your main speed issue is your datasource. Since you aren't writing direct you'd want to know how many hops it goes through and try to get some latency measurement there - and see if it degrades during burst periods. Second issue is going to be latency between when you send out an order and when it get's to your broker and then they forward it on to the exchange. You have distance, network latency, their internal latency and message queuing. For the most part you're not going to get a straight answer on any of that stuff. Which is more to the original point. Your own software latency is going to be pretty insignificant relative to everything else and really shouldn't be a consideration because the expense of getting the other latencies low enough to where it starts to matter. For IB in particular. Look for places to host and have them ping IB and give you the times. I believe it's in New Jersey. I don't know anything about retail alternatives to NxCore. For stocks, I've only used IB, NxCore, LIME and direct. LIME was decently fast but allowed you to short precisely nothing.
Thanks for taking the time with your answer - appreciated. As Nab mentioned above, IB's processing time is one of my main concerns. If their order receipt-to-arriving at exchange time is > 50ms, then it probably makes the project DOA. I wonder if there's any difference in processing time for FIX vs. regular API (wouldn't think so, but who knows). Assuming that's not an issue (and ignoring the Utah possibility for the moment), I believe the closest NxCore server to IB is in the Virginia Amazon Cloud. However, in order to access it I think you need to be set up there as well. I would guess that's a better alternative that colocating in NJ and waiting on data from Chicago (next closest NxCore server) - this has the obvious drawback of waiting longer to receive acks and such from IB, but for my strats that is less important that quickly reacting to market events. So if I went with that set up, I imagine exchange data to NxCore in VA is probably on the order of 10-15ms, add a couple ms to get to my Amazon instance, 1-2ms for strategy logic, plus 10-15ms more for my order to travel from VA to IB in NJ...a total of ~35ms plus IB processing time. Let me know if any of that sounds unreasonable - obviously the IB issue is likely the biggest wildcard...will have to think about some ways to test it under my current set up while isolating reporting time, network latency, etc.. If things look okay on the IB side, I can always check out other data feeds and colo in NJ...would guess that will cost 1-2k more monthly.