Example 25 – Other Ways to Do Predictions

NOTE: In NeuroShell Trader, open the chart named “Example 25 – Other Ways to Do Predictions” which is the basis for following example:

So far in our examples we have only showed you how to use the Prediction Wizard to build trading systems out of neural nets. The examples have been about making a prediction about the future, and then comparing that prediction to thresholds in order to be able to make a decision about whether to enter or exit long or short. In this example we will show you how to get the actual predictions out of the Prediction Wizard, and we explain the difference between the prediction data series and the prediction signal data series (that difference being something very hard for many people to understand initially).

So we start with a new premise – that we want to predict tomorrow’s high, and we want to see that predicted high plotted on the chart, not see trading signals based on it. The reason we may want that, for example, is to guide tomorrow’s discretionary day trading. The example does that, not with an eye towards getting a great prediction; we concentrate more on explaining the concepts. Before we discuss how we did that, let us first discuss how to read the results on the chart, which is what many people find hard to understand.

Note that the chart contains these four time series, which we will abbreviate AS, PS, A, and P for brevity:

AS – Actual Signal

PS – Predicted Signal

A – Actual

P – Prediction

In NeuroShell, we use the word “signal” to represent the prediction plotted on the day the prediction is made. We omit the word “signal” to represent the prediction plotted on the day for which the prediction is supposed to be true. If you look at the curves P and PS you see that they are exactly the same, except that P is shifted one bar to the right of PS.

Before we go further, we’d like you to click on the crosshair toolbar, the one that looks like you are looking through a rifle sight. A vertical line should appear on June 18, 2010 – if not click on June 18. Now locate in the historical view the values we are about to discuss (you might have to widen the columns using the mouse to slide the vertical column dividers until the full label for that column is visible):

On June 18 the Prediction Wizard predicted that the June 21 high (the next trading day) would be 14.11604. The correct June 21 high was 14.28. You can see these values in the columns for PS, AS and the Future value of Prediction for June 18. You can also see these same values in the columns for P and A for June 21.

You can use P or PS, whichever view of the prediction makes you the most comfortable. However, any time you use a neural net in a trading strategy, you must base decisions on PS, not on P, because the model won’t “see” P until it is too late!

Now look to or scroll to the right end (most recent part) of the chart, June 28. Note that PS does not go any farther on the chart than the last day on the chart. P goes one bar farther to June 29, the next trading day, which is yet to come.

Note that A goes to 6/28 also, but AS only goes to 6/25 (one trading day before).

Now we shall discuss how we made the prediction. If you modify (or just double click) one of P, PS, A, or AS and get back into the Prediction Wizard you can examine what we did. In the Positions tab you will see that we specified no trading positions. In the Training tab you will see that we specified an error objective function, rather than a profit one. And then in the Output tab, we chose the High as the neural net output. Those were the key ingredients needed to eliminate the trading signals and concentrate on the prediction itself.

Another thing which is necessary to get P and A as well as PS and AS is to turn on an option. Go to the Tools Menu -> Options -> Advanced and check the box called “List prediction data series”. You must do that before you train your network.

Frankly, it is more reasonable to predict the change or percent change in high than the high itself. That is because raw price levels are not normalized over time. However, the example was made with the prediction of the High because it is easier to explain that way.

Was this article helpful?

Related Articles