You would do that by defining your own "fitness function" and exposing
its value, for each backtest results row, as a named custom metric.
You would then use that named metric as the optimization target from the
walk forward settings dialog. From that point on, your fitness result
becomes the value which gets optimized.
In your case, your fitness function would be the value of CAR multiplied
by a (possibly sliding) scale factor taking into account the number of
trades.
You can see the idea in the code snippet below (which is using Ulcer
Performance Index rather than CAR). The code is just a snippet taken
from a larger implementation that included additional weights and had
custom position sizing on a bar by bar basis. So, the snippet you see is
not the most efficient in the context of this reply, but should give you
the idea. You may need to use a multiplier (e.g. 100 in example) to
raise values to larger numbers for display purposes only.
A few words of caution:
1. This will help guide the optimizer, so it will give less weight to
lesser values of your fitness function. But, it comes at the cost of
custom backtest calculations which are not free. It may end up taking
longer than without the guidance. But, it will result in values that
better suit your wants.
2. Keep in mind that you probably need a ratio of trades for the
period under study, as opposed to a fixed one size fits all number of
trades regardless of period duration.
3. The Optimizers (e.g. CMAE) are designed to work against continuous
fitness values. A simple on/off, high value/zero fitness will not work
well and can lead to very long optimizations (in my experience). Even
using the sliding scale of the example below, it would be better to have
a more continuous scale for when there are too few trades rather than
the sudden drop off factor of 0.001.
4. Be sure to type in the exact name of your custom metric, including
uppercase/lowercase as the target in the walk forward settings.
5. CAR is a notoriously poor estimator of fitness.
Mike
SetCustomBacktestProc("");
inRange = Status("barinrange");
barsInRange = 0;
if (Status("action") == actionPortfolio) {
bo = GetBacktesterObject();
bo.Backtest();
for (bar = 0; bar < BarCount; bar++) {
if (inRange[bar]) {
barsInRange++;
}
stats = bo.getPerformanceStats(0);
trades = stats.getValue("AllQty")/(barsInRange/252); // 252 bars per
year
if (trades <= 252) {
tMult = .001; // Penalize low trade volume
} else if (trades >= 504) {
tMult = 1; // 2 or more trades/day ideal
} else {
tMult = (trades - 252)/252; // Ratio of ideal
}
upi = stats.getValue("UlcerPerformanceIndex");
optimal = upi * 100 * tMult;
bo.addCustomMetric("Optimal", optimal);
bo.addCustomMetric("tMult", tMult);
}
--- In [email protected], Richard <richard.m.fe...@...> wrote:
>
> Is it possibile optimize with a target like CAR excluding results with
a
> number of trades to low?
> On this way the time required for optimization can be saved?
>
