Econometrics: State-space model forecasting - page 2

 
Integer:

Two options you can try: a prediction above the real value and a prediction above the previous prediction.
You need to be sure that the movement will be in the right direction and larger than the spread .
 
EconModel:
You need the certainty that the movement will be in the right direction and more spread .

What is the tester for? That's what it was invented for.

The trick here is different - your trick is a curvafitter, which automatically reduces the chances of the model's relevance to almost zero.

 
solar:

purple forecast, draft network of the old one.

bullshit...agree ?
 

EconModel, your result does not delight me. I got similar curves with ordinary nerve grids as well. But there were no profits. So they don't tell me much.

You can't go anywhere without advisor: if you just look at these "almost matching" curves and consider it a grail, you will not get any results.

1) Examine the difference between predictions and reality, and try (better - with statistical reasoning) experimentally to find a reasonable stop-loss and take-profit, so that the system will be profitable.

2. Another option: try predicting not one step ahead, but a few steps ahead. And then do the first point above.

P.S. I know who should be here soon... Will be even more fun.

 
Mathemat:

I've gotten similar curves with regular nerve grids as well.

the nets are easy on the head )))

siii-fact, yellow-forecast, oos after vertic black...


 
TheXpert:

And the tester for what? That's what it was invented for.

The trick here is different - your craft is a curvafitter, which automatically reduces the chances of the model's relevance to practically zero.


The word " curwafitter" is not familiar to me and google gave a backlink to this thread.

If you would be so kind as to state your thought in words known to me (or at least to google), I will certainly answer you.

 
solar:

I'm talking about how to use the forecast. I mean a man has made a forecast but does not know what to do with it. I've already written that this model works fine when the market is calm more or less. But I have long been concerned with searching for high and low )))). So personally for me, my version on the screenshot - a hirnya.
Characteristics of the model: non-linear for a non-stationary process. Adaptive: there are actually 18 AIC models used and a choice is made between them. But this choice is on a history of 48 bars?! The window size has to be reduced, but I don't know the criteria for choosing the window size.
 
EconModel:
The window size has to be reduced, but I don't know the criteria for choosing the window size.

You have to adjust it manually. Generally, it depends on the number of variables
 
Mathemat:

EconModel, your result does not delight me. I got similar curves with normal nerve-racking. But there were no profits. So they don't tell me much.

You can't go anywhere without an advisor: if you just look at these "almost matching" curves and think it's a grail, you won't get any results.

1. research the difference between prediction and reality and try (better - with statistical justification) to experimentally find a reasonable stop-loss and take-profit to make the system profitable.

2. Another option: try predicting not one step ahead, but a few steps ahead. And then do the first point above.

P.S. I know who should be here soon... will be even more fun.

The model is made on R. So there is a bunch of related information. Here is the MSE forecast error - Mean Square Error - RMS error. This is the square of the error. Below is the graph, but an error comparable to the eurusd level can be obtained by extracting the root.

I.e. for the maximum error the comparable one is 0.001871


Now I understand: it is necessary to enter, if the forecast exceeds those 18 pips? Or the average? Or is it a sko?

 
Demi:
manually select. In general, it depends on the number of variables

I think with a tester, but you need a criterion. It could well be minimal.