Message Font: Serif | Sans-Serif

No. of Recommendations: 3
What you say makes sense, but I'm getting the opposite conclusion from a GTR1 run

Your example is a 126 day hold, which is 6 months, not 6 weeks.

My point is that the turnover will of course go up when using a shorter
hold period within the "usual" range of 1-6 weeks, but it doesn't go up
nearly as much as you might think, because RRS slopes change slowly.

With 0.35% friction, 10 stocks, CAGR and annual turnover

Hold 5 31.4 / 5.72
Hold 10 31.2 / 4.89
Hold 15 31.1 / 4.45
Hold 20 31.3/ 4.16
Hold 25 30.8 / 3.93
Hold 30 30.4 / 3.73

Contrast that with a total return check tr(0,189):
Hold 5 17.4 / 11.66
Hold 30 26.5 / 4.27

In this case, the turnover rises rapidly and the friction eats any
possible improvement in the total returns.

So, the lesson is, if there is a way to calculate your relative
strength formula in a way that doesn't change the sort order too
rapidly, it can be worthwhile to have a system which checks frequently
but trades infrequently, which is what the RtW tends to do.

Frequency of tradnig can be brought down with any of the "usual" techniques,
including hold-till-drop or (as here using slopes rather than hard
total return periods. If I were using RtW, I would be tempted
to replace the prices 50 days ago and 225+50 days ago with smoothed
versions of those. So, for example, replace the price 50 days ago
with the average(48-52) days ago, and replace the price 275 days ago
with the average(270-280) days ago. This reduces the influence of
data errors, unusual temporary price movements in the past, and any
over sensitivity to the tuning valuse of the lookback, while also
reducing the frequency of changing of sort order a little bit. This
might reduce the need for HTD and conceivably improve returns a bit.
It's much more debatable whether the current price should be
replaced with (for example) the average over the last week, since you
want to catch downturns quickly. WMA(5) might be a nice compromise.
Note, I would probably do this smoothing even if it hurt a tiny bit
in backtest, since the backtest could be slightly less in danger of
overtuning to a particular pair of lookbacks.

Jim