Well, I think anyone here is familiar with slippage and commission, so I don't need to really explain those terms. When you test a strategy, you simulate trades under realistic conditions. That means you have a broker model that simulates a certain broker; the results would be different with any broker. The broker model should be managed and updated by the trade platform. Otherwise you'd be forced to script such a model with slippage or commission in your strategy, and change it anytime when your broker changes the spread. The results that I posted in this thread use the FXCM model, at a spread of 2.6 Pip. As FXCM is not a cheap broker, you will probably get slightly better results when simulating other brokers. But FXCM is the only model that I have at the moment. Most broker parameters such as rollover, commission, spread, are trivial or can be even zero, as in the case of commission. Slippage is not trivial and depends on the overall trade volume as well as on the price slope. It must be measured using a real money account for getting a slippage matrix that you then use in the test. Otherwise you have to estimate it. By the way, contrary to what I've seen on some trader websites, slippage is measured in seconds, not in percent or Pip or Dollars.
Or you could just look at the book, and see what you'll be matching with -- assuming you want to pay the spread. Edit: Is big slippage, measured in "seconds" good or bad? :eek:
Commission may be trivial to include, but the consequence is non-trivial and in fact in most cases dominating. Yes, estimating slippage is the key -- as in, run your back test, then run a live forward test, then compare the executions (your theoretical entry exits in your back test versus the actual entry exit prices you got). Once you do that you can histogram all those deltas and decide how you want to model them.
Good point: you could look at the spread and estimate the slippage based on some percentage of that spread at the time. The problem with this is in back testing you don't really have the book data so it's difficult or impossible to implement.
One way to stop debating and estimating slippage is to go get the book data, to see what you'll actually match with. If not, then at least be including the opposite best size in your backtest, to know whether you've broken through the top layer. Hope everyone out there backtesting away is doing this ..
Is historical intraday book data available? I would imagine it to be extremely expensive and would require enormous storage... It would have to be just snapshots of the book on the minutes timescale. I am sure the big players have thought of this and are already collecting.
Availability varies based on the exchange, and size varies according to .. the number of depth updates (obviously), which can be very instrument specific. For those who are cheap .. it's not real hard to capture a load of book data from, say IB, and as far as I'm aware, that's OK for personal use. May take some patience to do this, but a week's worth is probably enough to be useful.
Great ideas, thanks I might try it. I would think you may need to do this periodically as the spread will change. The other uncertainty is that a big determinant of the slippage is market orders (or lack thereof) for which we have no idea about.
Maybe. We know the minimum spread is bounded by a minimum tick size. And we know how to observe the spread. So, what's the distribution look like? Also there's a battle going on here. Passive limit orders in the book vs Market orders blundering across the spread to consume them. Who wins? Perhaps whoever's bigger. Or is that who loses? Time to build a model, eh?