Machine learning in trading: theory, models, practice and algo-trading - page 246

 
Awl Writer:

1)We train a neural network with N inputs on some area (input data cloud) in N-dimensional space. Neural network has no idea what to do with data that lie outside this cloud. But we feed it with this data and expect it to produce some result.



2) Dear experts, is there any mention of such things in the MO literature and does it seem useful to you? Comments are welcome.

1) this is called retraining, how to solve this problem is described in every literature on networks

2) I am NOT an expert, but I know how experts do it... it is better to immediately replace the price function with something else but very approximate with more suitable for you characteristics ... this is called a function approximation.

 
mytarmailS:

2) I am NOT an expert, but I know how experts do it... it is better to immediately replace the price function with something else but very approximate with more suitable for you characteristics ... this is called a function approximation

I've always said that the best predictor is MA. You don't need any networks. Everything is clear as it is. And the probability 0.5 just with MA is not a problem. It is clear, "if you add some butter to the wood", we can make it larger - 0.6-0.65. It is quite enough. And this is quite enough. Yes, by the way, even 0.5 is already very good.
 
Yuriy Asaulenko:
Yes, by the way, even 0.5 is already very good.
0.5 accidentally, it's a plum, even I know, what's good here?
 
pantural:
0.5 is random, it's a plum, even I know, so what's good here?

Are you sure? It's a flop at roulette. If at probability 0.5 you win 3-4 units, and lose one - it's a flop?

Read this forum, many profitable strategies work fine with probability 0.5.

FYI, in poker 1/6 and even 1/9 odds are good to win.)

 
Is it even possible to calculate that some time series in any way can be predicted or not?
 
pantural:
Is it even possible to calculate that some time series can be predicted in any way or not?
The best prediction is the current state. Considering that it will almost never stay that way.
 
Yuriy Asaulenko:

Are you sure? It's a flop at roulette. If at probability 0.5 you win 3-4 units, and lose one - it's a flop?

Read this forum, many profitable strategies work fine with 0.5 probability.

With a probability of 0.5 you'll be at zero on average, if you don't count transaction costs. 3-4 vs one is 70% (0.7) There are no winning strategies with 0.5 in principle.
 
pantural:
Is it even possible to calculate that some time series can be predicted in any way or not?
There are various tests for the degree of randomness, but not for the degree of predictability, as far as I know.
 
pantural:
With 0.5 probability you'll average zero if you don't count transaction costs. 3-4 against one is about 70% (0.7) There are no winning strategies with 0.5 in principle.
You're confusing the market with roulette or flipping a coin. The 0.5 probability is the probability of making the right entry, not the profit/loss ratio. And in poker, with a probability of winning 1/6-1/9 it's better not to sit down at all.))
 
Yuriy Asaulenko:
You are confusing the market with roulette or flipping a coin. The 0.5 probability is the probability of a correct entry, not the profit/loss ratio.
I'm talking about prediction statistics, how often you guess the direction of the market