Machine learning in trading: theory, models, practice and algo-trading - page 2659

New comment
 
mytarmailS #:

So the problem is not in pixels, or returns, but in our primitive models that we build for the market, the maximum of these models is to multiply pixel by pixel and look at it in a sliding window of 5-10 pixels) that's the whole fairy tale. That is, the model does not rise above the first level of abstraction, and we may need 1000 such levels....


So don't scold returns or pixels, you need to think more with your head, and of course you need knowledge....

Well, good luck to you in mastering new levels. I sincerely wish that you succeed, although I doubt it very much.

 
Valeriy Yastremskiy #:
When you turn off the gas on sizzling cutlets, why do they keep sizzling. Because of inertia. )))) and that's the correct answer.
))
 
sibirqk #:
If someone could learn to predict at the opening of each minute whether the price will be higher or lower in 240 minutes (i.e. the colour of a 4-hour candle), with a probability of at least 55/45 - that would be alpha.

This is already available https://www.mql5.com/ru/blogs/post/746398
Only at every minute it makes no sense to forecast for 4 hours ahead. Too smeared forecast, it turns out 240 forecasts inside one duration. I.e. it is possible, but it is an inefficient use of the resource.

Forecast for any duration is built from minute candles, but the poll itself is optimal to make every 1/10 of the forecast range, i.e. 4 hours is normal to forecast on 20 minute candles (when closing 20 minute candle to start the poll). Otherwise there will be a lot of similar answers, because when shifting to 1/240 nothing fundamentally changes.

As for the quality of the forecast, it turned out to depend on the trading pair and the duration of the forecast. Crypto and large indices like SNP500 are best predicted, while commodities and everything that strongly depends on news and politics are the worst (indices depend, but more smoothly). Bitcoin now has 68% successful predictions on a 3 hour duration, the best rate there is. This figure is derived from real public forecasts on the real market.

On average, you can get around 60% for many pairs and durations from 7 minutes to 6 hours.

That is, the level of a good indicator has already been reached.

 
Evgeny Dyuka #:

This is already there https://www.mql5.com/ru/blogs/post/746398
Only at every minute it makes no sense to forecast 4 hours ahead. Too smeared forecast, it turns out to be 240 forecasts within one duration. I.e. it is possible, but it is an inefficient use of the resource.

Forecast for any duration is built from minute candles, but the survey itself is optimal to make every 1/10 of the forecast range, i.e. 4 hours is normal to forecast on 20 minute candles (when closing 20 minute candle start the survey). Otherwise there will be a lot of similar answers, because nothing changes fundamentally when shifting to 1/240.

The specifics of implementation are naturally selected from real trading conditions and opportunities. I was illustrating the principle itself. I did some modelling - the EUR/USD open minutes were read from a file and with a given lag (e.g. 240 min) the profit in pips was calculated. With a given probability it could be positive or negative and the spread was immediately deducted. The spread was taken to take into account any surprises of dealing. The history was taken for five years and chased in a cycle for thousands and fifty times to get the expectation profit on each combination of parameters. It turned out that when the probability increases to 55/45, the spread has little effect on the total profit, and the final result is quite decent.



 
Evgeny Dyuka #:


As for the quality of the forecast, as it turned out, it depends on the trading pair and the duration of the forecast. Crypto and large indices like SNP500 are the best predicted, while commodities and everything that is highly dependent on news and politics are the worst (indices are dependent, but in a smoother way). Bitcoin now has 68% successful predictions on a 3 hour duration, the best rate there is. This figure is derived from real public forecasts on the real market.

On average, you can get around 60% for many pairs and durations from 7 minutes to 6 hours.

I.e. the level of a good indicator has already been reached.

In my opinion, it is necessary to take into account not only the quality of the forecast, but also the influence of the spread. On small intervals it can be higher, but the influence of the spread on the final result is stronger.
 
Valeriy Yastremskiy #:
When you turn off the gas on sizzling cutlets, why do they keep sizzling. Because of inertia. )))) and that is the correct answer.
The physics of financial markets and the physics of the real world are two big differences. The physics of financial markets, in my opinion, is closer to the children's toy kaleidoscope than to cutlets in a frying pan. With each new turn (time step) the combination of colours (market conditions) changes dramatically. Of course, there is inertia, but it is difficult to identify it and it is not as unambiguous as that of the cutlets on the frying pan.
 
Maxim Dmitrievsky #:
And why is algotrading not 100% in the signal, how does that even work?
 
mytarmailS #:
And why in the signal algotrading is not 100%, how does it even work?

are you talking about the demo? i don't know, maybe i closed something with my hands before, that account is for tests, don't pay attention.

 
Maxim Dmitrievsky #:

You mean the demo? I don't know, maybe I closed something with my hands before, that account is for tests, don't pay attention.

Yes, about the demo...
I'm just curious and not clear how their algorithm determines whether the trading is manual or algorithmic.
 
mytarmailS #:
Yeah, about the demo...
I'm just curious and not clear how their algorithm determines whether the trading is manual or algorithmic.

through the trade IDs, there's a marker.

1...265226532654265526562657265826592660266126622663266426652666...3574
New comment