I pretty much wrote all my current thinking on de-trending in previous posts and comments, but I’ll try to recap. The two previous mentions here were more on systems blends and low de-trended correlation and benefits of diversification, especially in the comments.

From what I’ve played with, finding systems with similar cumulative annualized growth rates (CAGR) is more important than finding systems with low de-trended correlation in their equity curves. If system A grows at 25% and system B grows at 5%, it doesn’t matter what the correlation is, a combination of those systems will probably be worse, in most risk-adjusted metrics, than system A alone. The few times where it may be better would be extremes of low correlation (negative in de-trended), combined with a risk-adjusted metric that had a low hurdle, like Sharpe or Sortino at risk-free rates, perhaps.

I think that de-trending has more importance when working with equity curves than it does when working with return series data. For example, 1.00, 1.05, 1.10, 1.05 can be an equity curve that looks a lot like 0%, 5%, 5%, -5% as a return series. With any two systems that have positive returns over time, the return series may or may not have a trend, but the equity curves most certainly will, since they will both be increasing over time (correlation with time).

There’s a lot of snake oil and bone-tossing in the field of finding and defining trends, which makes removing trend from a series all the more one-off and customized. Is the trend in a time series logarithmic, exponential, linear, a power series? Is there one trend or more than one, occurring at different points in the series? Any good intro stats link, like the ones on my links page, will have information on linear regression, nonlinear estimation, and data transforms.

I don’t have any collection of links to outside articles on the process, although a Scholar search or “advanced” using file type *.PDF at Google may get you a list. In my opinion, the only way to get a feel for what de-trended series look like and taste like is to download some series, get you hands dirty estimating the trends, and then “remove” them and see what the series look and taste like.

John recommended Hodrick-Prescott, which can be found in a Google search. It seems to me to be a cookbook method that assumes economic data breaks cleanly into “logarithmic trend” + “cyclical” + “approximately normally-distributed noise” components. Maybe economic data does (and maybe it doesn’t), but I don’t think that applies to a back-tested equity curve from a trading system.

Bottom line is get your hands dirty in a spreadsheet, de-trending data. Try different estimating processes for the trend, linear or logarithmic, check for different trends at different periods, and run some correlation studies before and after the de-trending. The SPDR series of sector ETFs are a good place to start, with many series running back almost a decade.