With all due respect, the OP is telling us his backtest covers 49,000 trades over 40 pairs, and that's all the feedback you can provide? I challenge you to come with any system, overfit at will, showing positive expectancy over 49,000 or more trades, using conservative assumptions for commissions & slippage. Please provide your tradelist as a proof.
1. You should normalize everything to a standard position-size 2. Ignore the backtest & live stdev 3. Using the backtest trades distribution, run MC sim for the number of live trades you currently have, and take that stdev 4. Divide your normalized live avg/trade by the stdev just computed : that's how far your current live avg/trade is vs. what you expect it to be based on the backtest. Given the very small live sample-size, the stdev should be sky-high, and the current distance as a result should be close to zero - that's my point : with such a small sample size, you can't conclude anything, unless your live results are really way far off. In the long run, I would consider live-results to be "as expected" as long as that distance is less than 2 stdev. (at some point, you will want to use the last N months of live trade data for this - even though lower sample size means higher stdev, it is also faster to detect a degradation into "unexpected" territory). As for your current live slippage being better than the backtest assumption, that's a good thing & I would not change the backtest assumptions - my experience is that once in a while, you experience a large unfavorable slippage which will eat a lot of the prior small favorable slippage.
dom, thanks a lot for your insight, highly appreciated. I may keep this thread updated weekly with live results until my system falls below the 2 std devs
I've backtested a lot of break-out systems and they usually show 60%+ winners. The only problem is: you usually get in with a buy or sell STOP order....so the problem is getting filled at a reasonable price when the instrument makes the break-out.
Why not just purchase some historic data and backtest this system with a couple dozen pairs ? This is much faster than live trading....which might take YEARS before you can make any conclusions.
You may have missed the beginning of this thread. This is already backtested on ~ 49.000 trades on historical data. It only works on a subset of pairs though, hence only those are what is being used for live trading.
It is still running. The overall curve did not flatten (much) - still within the same parameters, however it doesn't absorb any size. The second I tried to significantly increase the size (10x) I ran into fill issues even though those are fairly liquid stocks.
Overall sounds like a win, but I hear you on the size/fills issue. I swear if I ever end up using an entity, I'm going to call it Not Very Scalable LLC. We may have similar set ups - I've been running automated strats with IB for a while...one of them seems similar to your original pairs strat. FWIW, I discovered something that I found rather counter intuitive, but that model ended up doing way better using LMT+MKT (less liquid leg first) rather than REL+MKT. Unfortunately, it was so long ago that I don't remember what it was about the execution mechanics that made it that way. Anyhow, just some food for thought. I'll feel terrible if you experiment and it's costly, so let's just agree to have Baron pick up the tab on any R&D losses. I've asked this question of some stat arb guys before but never really heard a satisfactory answer...curious if you have a take: What rationale did you settle on for a related pair (say sector peers) to trend rather than mean revert? Feel free to pass if that's too close to any secret sauce - really just interested in a for dummies answer as it's not at all obvious to me.