Very simple about Back Testing


I think that the best way to discuss Back Testing in brief is trying something that TS users do every day. As an example, we may start with the question; "How do we know that we can trust this projection line based on Jupiter cycle"?


I hope that you all know how to get this projection line. Let's look for the answer to the question above together. Before we start, I have one condition: you need really understand what we do. No special knowledge is required - just a common sense.

The simplest answer to the question is: wait 10-20 years, you will know for sure whether this forecast is good or not. I believe everything is clear here.

So we can imagine this situation: the forecast based on Jupiter cycle has been made 20 years ago, in the year 1989. A person who made it did what we all know: he downloaded all available at that time price history for Dow (since 1885 till 1989) and got this forecast::

Why do we conduct this experiment? The answer is very simple: now, in the year 2009, we know what happened to Dow after the year 1989, so we can see how that forecast really worked. Here it is:


This is the essence of the Back Testing approach that I use. Here we assume that if some model worked well enough in the past, it might work rather well now and for the future. 

How to make this experiment with Timing Solution?

It is practically automated: you need to do just one mouse click, to put Leaning Border Cursor (LBC) to the year 1989, and the program will recalculate the composite diagram for Jupiter cycle:


The LBC is the border between the past (this is what we know) and the future (that we do not know yet). In other words, the program "does not  know" the price movement after LBC. We can conduct this imaginary experiment  with one mouse click and see the forecasting ability of the analyzed cycle, just imagine yourself investing your own money in 1989 based on this projection line.

The more chances we have to test our model on different time intervals, the more we can trust it. You can try it moving LBC to different positions.


There is one more possibility here. Clicking this button, you will make invisible the price history beyond LBC:

This is how it looks now:


Thus you can push this button and conduct your analysis, then release this button and see how your models fit the real price movement. 

Even now you see that it is time consuming. Now imagine how much time it will take when you use more complicated models, not just Jupiter cycle. And you can vary there the parameters of those models. And test all modifications on many time intervals... Have you imagined that? Now you see why we need to use Neural Nets for Back Testing - just to be able to deal with this sea of information. Nothing more.