Hi All,I would like to know how one can do a proper back-test a HFT strategy. I already reconstructed the order book with milliseconds time stamps so I have 10 (could be as many as I want) market levels at any given time during the day. There are 1.3 1.5 million order entries (new orders, updates, cancelling and trade) during one day in this security.Also, in the local market using a collocation server a round trip for 1 order takes 1.3 milliseconds. So my question is how can I presume that my quotes would be filled : Receive the signal + 0.65 milliseconds and see best quote out there? How much should I add toprocess the signal and how much more to the computer to process the live data feed? Thanks !

This may help you:http://papers.ssrn.com/sol3/papers.cfm? ... id=2308659

There are many approaches to this problem, and none of them can be successful without taking into account the competition. If there is no competition for the opportunity you are going after, then you can assume 100% fill rate. It gets interesting when you are going after the same opportunities with others, which you can find out by looking at the trades that occur after you send the order. In such a scenario, you need to get distributions on your success rates with respect to latency. This will give an indication of where you stand within competition, if you're slow, you're more likely to have less success. On top of the pure performance comparison, you need to take into account the variance of the exchange systems, which can be measured independently. The combination of the two will give you a good start for simulation. In general you can model any queue with a poisson distribution, so having long tails. So if you extract the parameters, you can even come up with an analytical solution.

How does one beat the rock-paper-scissors robot? With a brick.

QuoteOriginally posted by: taylanThere are many approaches to this problem, and none of them can be successful without taking into account the competition. If there is no competition for the opportunity you are going after, then you can assume 100% fill rate. It gets interesting when you are going after the same opportunities with others, which you can find out by looking at the trades that occur after you send the order. In such a scenario, you need to get distributions on your success rates with respect to latency. This will give an indication of where you stand within competition, if you're slow, you're more likely to have less success. On top of the pure performance comparison, you need to take into account the variance of the exchange systems, which can be measured independently. The combination of the two will give you a good start for simulation. In general you can model any queue with a poisson distribution, so having long tails. So if you extract the parameters, you can even come up with an analytical solution.These are a good points actually. So let me see if I got this right. Instead of measuring the P&L, first take a look at the orders that happen after the signal was sent? I was thinking one step before that, how to measure my tick-to-order latency but this surely helped.