jebus89 wrote:

and years later its obvious that you should have let your system be as it is..

Or even weeks / months later? 🙂

I often have 4 to 6 versions of the same System on Forward Test and it is not often that the forked / improved versions giver continued better results than the original.  But the few that do (give better results) make the effort worthwhile.

I look for improvements on all fronts … Gain to Loss Ratio, % of winning trades, drawdown, etc.  I would settle for much improved profit at the expense of the other stats … providing trades enter and exit at logical price action points.

What spooks me is profits due to trades sitting there for ever and ever – with big drawdown – but coming up good in the end … good luck which likely will turn bad very soon!?

I always look at what can be taken away to make a strategy less curve fitted rather than what I can add which will only ever make it more curve fitted.

I totally get what you say about not wanting to hold on too long to negative trades. Holding is fine if you start with a big bank but then if you have a big bank then just buy stocks and hold on to them through thick and thin – don’t spread bet! I would rather look for a lot of strategies that are very simple and have a slight edge and just buy today and sell tomorrow or buy today and sell in a weeks time. Trade a lot of these uncorrelated and you will make a profit – but the important thing is to keep it very simple – less is more when it comes to conditions in a strategy in my humble opinion.

I think that you always have to justify every condition by analysing it before you allow it to stay in a strategy.

Take the humble average as a long or short filter for example. I would want to test the strategy with every average period of a large range and then plot the results in Excel charts. I would want to see a pretty large area in the chart where the results are not too dissimilar in returns and win rates. This means we now have a large range of average periods that have worked pretty well historically even if not all the equity curves would be to our liking. I would then choose a period for our average that is smack bang in the middle of this working range of periods or the one with our nicest results if it is near the middle of the range which should minimise our new filter being stood on a cliff edge.

You could even consider running three or four identical strategies but just with different average periods from across this range of working periods and then average out the equity curve.

If we add a condition that gives a more binary result – i.e. it works with a certain setting but does not work with any other setting then I really would want to delve deep and see if I can understand why it works. This is where robustness testing is useful – especially if done on OOS data.