Whether there is a process whose analysis of one part does not allow predicting the next part. - page 13

New comment
 
joo:

5. Genetics of its own making.


Actually, I don't believe it is possible to predict random wandering. Another question is whether it is possible to create a trading strategy that would be profitable at least for short periods of time. The answer is more likely if we use process statistics to our advantage. Martingale, for example.

Your genetic algorithm is powerful if it has managed to unlock the random number generation algorithm. Will you be betting on it for the championship? I'd like to cheer for it. Although the conditions of acceptance (15 min testing) are tough. I am thinking of using my old Expert Advisor based on the nearest neighbour search. But it will be difficult for him to fit in 15 minutes. Championship conditions are not designed for neural networks.

 
Random walk cannot be predicted. You can predict something that has regularities, there are no regularities in random walk by definition. If you run more data through this network, and if the series is indeed random, the profit will go to zero, the probability of prediction will be 0.5.
 
Avals:
Of course NS can memorize among other random data. Is it a prediction?

You're right, it can and does remember.

My experiment was incorrect - Sample and OOS were generated at different times, so they were not random numbers from the same row (the generator was initialised differently in both cases).

But here is a correct experiment:

A sequence of 10000 OOS Sample and immediately followed by 10000 OOS OOS were generated.

Is it a prediction now?

 
gpwr:


1. Actually I don't believe it's possible to predict a random stray. Another question is whether it is possible to create a trading strategy that would be profitable at least for short periods of time. The answer is more likely if we use process statistics to our advantage. Martingale, for example.

2. Your genetic algorithm is powerful if it manages to unlock the random number generation algorithm. Will you be betting on him for the championship? I'd like to cheer for it. Although the conditions of acceptance (15 min testing) are tough. I am thinking of using my old EA based on the nearest neighbour search. But it will be difficult for him to fit in 15 minutes. Championship conditions are not created for neural networks.

1. I don't believe it either. :) I only believe that on real market data the results should be even better than on PCF.

2. Yes, GA is really good, heck - not for nothing I spent a total of about 3 years on its development. I'll probably do it this year (whew, whew, whew), and as for the grid limit during testing, I should firmly prescribe ready answers for the grid right in the EA so it doesn't "think" during testing.

 
Integer:
Random wandering and you can't predict. You can predict something that has a pattern...

I think so too.

 
alsu:
By the way, in terms of performance, MathRand()&0x00000001 is better

You can't do that with a C PRNG. It has a "dirty" loop, so the result of such an operation will have cyclic patterns.


ZS, it turns out that we've already corrected it.

 
The moral of this thread is that the NS can remember even the SB, but she cannot make an SB prediction.
 

And this is already on the real data, EURUSD M5. Sample 2011.11.01-2012.05.25, 10958 samples green curve, and Backward 2011.09.01-2011.11.01, 5160 samples red curve.

Well, it's just a saying.

 
Backward is the magic word )
 
TheXpert:
Backward is the magic word )

Forward is from Sample

Backward - backward from Sample

1...67891011121314151617
New comment