Which price curve are you using for an automated trading system? The close price is the most recent, while (H+L)/2 lags by half a bar but has less noise. Has anyone experienced with the effects of different prices on automated trade results?

You cannot enter a trade at the next bar at (H+L)/2 but you can enter on close. Thus, use of this series may invoke forward looking.

You're right of course, but I do not mean the price of entering a trade. I meant the price series used for trade decisions, f.i. for moving averages, high pass filters and so on. The price of entering a trade at market is of course the open price of the next bar plus the slippage.

Depends on the system and on the time aggregation. At aggregations > 5 min, (H+L+C)/3 I have found to be a pretty decent way of finding the direction of the market, for trend following systems. Never found (H+L)/2 to be useful.

I found that (H+L+C)/3 gives annual returns between C and (H+L)/2, with a multi asset strategy with 4 hour bars: (H+L)/2 - 252%, error 14% (H+L+C)/3 - 234%, error 17% C - 221%, error 20% So all are pretty close, but I only have this strategy so far. I'd be interested in results with different strategies.

Hi jcl, what is "error" in this table? Is it the same as sigma? What is the standard deviation of daily log returns using the (H+L)/2 method for your strat?

No, the error term only measures the influence of random noise on the result. For calculating it, the result of the original price curve is compared with artificial price curves where the bars are constructed with a tick offset. This gives some information about the reliability of the result. I don't have the standard deviations of the log returns though - I only calculate the standard deviation of the per-bar returns for getting the Sharpe ratio. The Sharpe ratio is about 2.1 for the three tests. For what do you use the standard deviation of the daily log returns?