Market etiquette or good manners in a minefield - page 34
You are missing trading opportunities:
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
Registration
Log in
You agree to website policy and terms of use
If you do not have an account, please register
to Neutron
Sergey, here is a graph of what I got in Matkad. Here, the red line is the first difference of the five-member sine, and the green line is the neuron's attempts to guess its behavior in the next step.
It seems to be working.
I will now do the AUDUSD minutes in Matcad. Oh, I forgot to mention that it was a neuron with binary inputs, and the picture below is the same neuron with actual inputs
Now I'll do the AUDUSD minutes in Matcad.
Hold on.
Let's present the results of your NS on the model series in a more informative way. For the NS with real inputs we will plot the predicted value of BP increment taking into account the sign and amplitude on abscissa, and the predicted value - on ordinate. With a sufficient number of experimental points, we will end up with the following picture:
Here, the lilac dots show the training sample, the blue dots show the test sample, and the black dots show the test sample on a random VR. Straight lines are drawn through the data clouds using the method of least squares. The angle of slope of such a line shows accuracy of the forecast (the closer it is to 45 degrees, the more accurate the forecast), spread of points around the line - forecast variance, the difference between the angle of slope of the black line and the horizontal direction - statistical significance of the obtained results and at the same time absence of algorithm errors (for example, looking into the future).
For a NS working with binary inputs and used to predict only the sign of expected movement, we can evaluate the prediction accuracy with just one parameter - the % of exact hits. It can be calculated by the following formula:
Where, x[i] is the real increment value, y[i] is the forecast value. This procedure should be used both for the training and for the test samples, it's better to have n more than 100.
Doing. Please clarify about the least squares method - I don't know what it is or how to calculate it. Oh, and one more thing: how do you paste images and formulas from Matkad into the forum?
I first copy them into a graphical editor and then crop them as needed.
Yes, the same way. I have a Screen Capture function in my graphics editor that captures the highlighted area of the monitor. You can also click on the graph and copy from the context menu, etc.
Here's the formula for the MOOC:
You need the x and y vector lengths to be the same. The method allows you to draw a line through a set of points such that the sum of all distances from each point to that line is the minimum of all possible choices.
I'm not getting a cloud like the one in your picture
I get this one:
Here on the abscissa axis is the first difference of the five-member sine, on the ordinate axis is the network forecast
Maybe it's the wrong type of graph? It tries to connect all points with lines. I'll try something else now
So, in the graphics settings, select dot representation instead of dotted representation:
and colour those dots the way you want them to be. And display the grid (settings in the window to the left).
О! A hysteresis loop is emerging, though!
Sergey, isn't training sample and test sample the same thing?
My neuron learns each time on n data vector samples, and predicts n+1th sample (of the same data vector). The difference between the prediction of the grid and the actual value of the n+1th sample I use to calculate the correction. Under these conditions, how can I plot the graphs for the training and test samples separately?
The method allows you to draw a line through a set of points such that the sum of all the distances from each point to that line is the minimum possible.
Let me correct you. Not "distances", but "squares of distances".
And, by the way, not only the coefficients of lines can be calculated by ANC.
Sergey, isn't training sample and test sample the same thing?
My neuron learns each time on n data vector samples, and predicts n+1th sample (of the same data vector). The difference between the prediction of the grid and the actual value of the n+1th sample I use to calculate the correction. Under such conditions, how to plot the graphs for the training and test samples separately?
It doesn't matter how we define them, what is important is that there are samples on which the NS was trained and which were predicted. That's how you get two rows of points.
Let me correct you. Not "distances", but "squares of distances".
And, by the way, it's not just straight line coefficients that can be calculated using ANC.