|
PureBytes Links
Trading Reference Links
|
Following the correspondence concerning Steve's system (Re: "Message
Interruptus"), this is a good time to warn any recent arrivals on the
scene of the dangers of accepting optimised results as reality. In a
system test with optimisation the computer selects parameters to give
the best possible results over the data tested. This is with the
benefit of hindsight. Future data will have different characteristics
and different parameters will apply. You cannot apply optimised
parameters to future data and expect the same results as in the test
data.
The way to use optimisation is to find an area with somewhat less than
the maximum profit, but where a range of nearby parameters produce
similar results. This indicates robustness: that the characteristics of
the data can vary but the parameters are still likely to give acceptable
results.
If sufficient data is available (and it never is sufficient) divide it
into two portions; optimise on one portion then apply the optimised
parameters to the other portion to see how the results compare. For
statistical significance there should be at least thirty trades in each
portion. Another way is Walk-forward Testing: test over say the first
five years and apply the optimised parameters to the sixth year. Then
move forward one year and repeat. Contimue to the end of the data,
keeping note of the parameters on each repeat. Then select parameters
in the middle of the range.
Perhaps others have comments to add?
Harvey Pearce, Victoria, B.C., Canada
=====================================
|