dcengr, here's a little more than a candlestick system you can trust if you want to do historical scans;
1. Calculate the width of the bollinger bands at multiple time frames, let's say 2.5, 5, 10, 20, 40, 80 wks for 2.5, 5, 10, 20, 40, 80 wk periods.
2. Calculate the CCIs of the price at these multiple time frames. So, you will have 6 time series with different lengths in each.
3. Find the closest matching patterns historically starting from the largest to the smallest, so first find the highest correlating 80 wk, then 40 within it etc. Iterate until the correlation falls below 60% in each of the time frames.
5. Repeat these until the 80wk correlations in the historical data falls below the 60% at least. Now, you have separated all of the possible time frames that might be relevant to the current time frame.
6. Calculate the BB widths and the CCIs for N/2 time periods in the separated time frames, so the BBs and CCIs in the following 40wk for the 80wk time periods, the following 20 wk for the 40 wk time periods etc...
7. Calculate the average expected bollinger band widths and the CCIs by weighting with the correlations in each of these N/2 time periods previously calculated...
8. Inverse transform the current prices according to your expected BB and CCI widths and make a forecast...
This is a pure statistical method by using the historical data. The momentum and cyclical data will allow you to mold the existing data and make a forecast without these since you evolve the determinant characteristics of the current prices, but you have to have a very good model. In fact, the better your model is the more accurate your historical scans will be. BTW, to be even more accurate, you can start with the currency correlations, then yields and then stock indices of the same time periods, or simply do a multiple correlation of all in each of the time periods above.
This is the kind of stuff I like, things that can be shown. But to be statistically relevant, more samples should be present. A test going back to 1900s may reveal something completely different.
A better way to test the signals is to test for the past 20 yrs without the past 3 yrs and then see whether the method you found by using the prior 20 yrs is still applicable to the past 3. This way you can also tell whether the methods are not outdated or discovered by too many...
- kisa