Verification of the projection line with Timing Solution
In brief, the rule to verify the projection line is this: "Watch how this projection line forecasts the future on the TESTING interval and ask yourself "can this projection line help you with your trading?". If this testing interval is too long, it might be very hard to estimate this projection line visually; you can calculate a correlation between the projection line and the initial data on the TESTING interval. Correlation starting from 10% is good enough. Also, remember that you need to have 3 years of the price history on your TESTING interval, better 5 years". This rule works for daily data and may be used for weekly data as well.
I formed this rule as simple as possible. However, it was not an easy task. Quite opposite, this is an extremely hard task to verify a projection line correctly, it is the most complicated and time consuming task I have ever met in my life. I know that for sure. Now, in "Terra Incognita" project, I use different approaches to this task in Trading Strategy Constructor and Turbo Cycles Backtesting modules. Before that I have tried to solve this problem with Neural Network Backtesting module, this is a big story ...
In this situation when the problem is very complicated I would recommend to use your common sense first. You have to see/feel this projection line, and only if you see that this projection works ok you may apply the formal criteria. Not the opposite: do not try to apply the formal criteria without preliminary visual estimation. I worked with financial data long enough and know well enough that all formal parameters - like win/loss, profit factor, equity curve - can easily give you an illusion of a good model. As I see, there are two ways here to handle this problem: use your common sense or gather a very good understanding of the nuances of financial mathematics (and still, use your common sense).
Now, how to do this visual analysis? First of all, you need to set Learning Border Cursor (LBC) so at least one year of the price history is left for the TESTING interval (1 or 3 years; 3 years is a must if you plan to use the correlation). The explanation regarding this issue is here: http://www.timingsolution.com/TS/Mini/39/index.htm\
Follow these steps:
Now run any module, suppose it is Neural Network module. It has provided us with the projection line like this:
Watch how this projection line works on the TESTING interval (it is out of sample interval). The movement of the projection line within this TESTING interval shows us the real workability of our model. The projection line prior LBC will always be good as this is a result of a normally working optimization procedure and it does not show the real workability of this projection line. Simply watch how this projection line fits the price on the TESTING interval, how well (or not well) it follows ups and downs of the price data..
If you have enough price history on the TESTING interval (3 years and more), you can calculate the correlation between the projection line and the price data on the testing interval. Follow these steps:
Thus you get the correlation between the price and the projection line within the TESTING interval.
Beware of the random error that can be very high here. Bette watch how this correlation is changing during the training procedure.