Training and Testing intervals. Learning Border Cursor (LBC)
When you download any price data file, you will see that all available data are divided into two intervals:
Blue interval is Training Interval, Red interval is Testing Interval. The border between them is Learning Border Cursor - LBC.
You can set the LBC wherever you want. To do this, just click on this icon: and then click the left mouse button at the place where you want to set LBC.
Also you can set the LBC clicking RIGHT mouse button and choose this item:
Also you can set LBC clicking on this button: . In this case, you will get the window to set LBC on any price point manually:
Back Testing Concept (how to avoid "information leaks")
When you create any forecasting model, the program takes the points from TRAINING interval only. It does not use the price points from TESTING interval, these points serve to estimate the model’s performance only.
Look at this example:
This is one of the models to forecast Dow Jones Industrial (DJI) index. This model is based on fixed cycles plus Neural Net. The Black line is the oscillator for DJI, the red line is the forecasting curve calculated by the Neural Net. It shows good fitness on the blue (training) interval. This is not surprising, because the program uses these points to train our Neural Net.
The fitness on red (testing) interval of the example is not too good. Here the Neural Net works in the forecasting regime. The price points on the testing interval are necessary to evaluate the real forecasting ability of the applied Neural Net model. Here we compare two curves - the data line and the projection line.
So, remember this: the real forecasting ability of any model can be estimated on the TESTING interval only. Thus we can avoid so called "information leaks".
The Neural Net knows and "sees" only the price data before the Learning Border Cursor when it creates the models.
For some tasks, it is necessary to use another interval; it is called VALIDATE INTERVAL, you can define the length of the validate interval here (in “Main Window View”):
Now the Main window looks:
It is divided on three intervals A – training, B- validate and C – testing. The ABC notification is used everywhere in the program.
For example, in Astronomy (Composite) module this means:
the correlation between projection line and price calculated on TESTING (C) interval is 0.1259.
Final optimization is when you use ALL available price points to make the forecasting model. When we do this, there is no price points left that can serve as a testing interval to do Back Testing, because now the program uses all of the price points to create the final model. Thus, final optimization is recommended only when you trust your model.
To make the final optimization, move the LBC to the last available price bar (by clicking on this icon ).
Here is how the forecasting model looks like:
Compare this picture to the diagram above. There is no price data on the red interval, when you do final optimization. You have here only real forecast.