Bayesian regression - Has anyone made an EA using this algorithm? - page 25
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
Thank you very much, sensei, for how generously you have shared, with me, an inexperienced, precious grains of truth ...
An "expert" showed up...
It's what I do for a living, so I don't give a shit about the attitudes of theorists, especially incompetent ones.)
And the most important question - is there a Result ? :)
Check.
Specifically about the model - out-of-sample prediction error is around 30%. The best is ada. Not much worse than random forest and SVM. But to get such a result I first had to learn how to select predictors. Generally 70% is spent on preparation of predictors (data mining), 10% on model and the rest on performance estimation. But it's not an Expert Advisor - it's just a model, but a model that doesn't have the property of overtraining and doesn't depend on stationarity or non-stationarity of the input data.
Once again trying to read something on the subject, I come to the conclusion - what a nightmare.
So it all starts with stationarity. What is the correct definition of stationarity?
I found this one:
Стационарность — свойство процесса не менять свои характеристики со временем.
What specific characteristics do you mean?
I keep reading:
Stationarityof a random process means that its probability regularities remain constant in time, and we usually consider two types of stationarity: stationarity in the narrow sense when finite-dimensional distributions are invariant to time shifts and stationarity in the broad sense when onlymathematical expectationsdo not depend on time. The practical application of stationarity is based on the fact that for a stationary process the characteristics of any random sample and the general population coincide.
What's that? -"when finite-dimensional distributions are invariant with respect to time shifts". What kind of finite-dimensional distributions are these? And invariant, is it supposed to be constant, i.e. do not change? And what do you mean by stationarity?
In a broad sense: time-independent mathematical expectation. So? Obviously by this definition market data is not stationary. And what is the point of talking about stationary data (of course with this definition). With such data it's elementary to make a profit, open towards mathematical expectation when it declines... Except that no one made a constant expectation here.
---
Here's the trick: We flip a coin. We record the results in two ways:
1. Heads +1. Tails -1. We have something like this: 1,1,-1,1,1,-1,-1,-1,-1,1,1.
2. Add it up. We have something like this: 0,1,2,1,2,3,2,1,0,1,2.
In the first case, expectation is constant, while in the second case it is not. So, in the first case the data is stationary, and in the second case it is non-stationary. So, like in the first case you have the possibility to gain profit and in the second case you don't? And neither in the first nor in the second there is a possibility to take a profit.
So what's the point of all this fuss about stationarity?
I come to the conclusion that stationarity is an indicator of nothing.
I am coming to the conclusion that stationarity is an indicator of nothing.
Now let's imagine for a second that the first row could be traded...
You can imagine anything. Suppose we open on the deviation towards the expectation. We have 1, sell. The next value is again 1, again 1, 1, 1, 1, 1, 1. Finally, 0 and -1. It would seem that this is a profit. But in fact the price has moved upwards and with the number -1, there is no profit. It turns out that the oscillator is as good as the oscillator.
In this case it wasn't necessary to analyze, because initially a coin was offered and no matter how you turn it, you won't make profit.
OK, the traded instrument is the difference between two futures of different delivery dates. Is that clearer? Yes, it is unrealistic because the fluctuations have a very narrow bell.
But if the traded instrument is the difference between a CME futures and a moex futures, the bell is wider and this is quite a real profit for advanced high-frequency traders.
Even more down-to-earth is the EURCHF in the period when the Swiss central bank pegged the franc to the euro. Only the laziest man hasn't traded that thing.