Machine learning in trading: theory, models, practice and algo-trading - page 3577

[Maxim Dmitrievsky]

Forester #:

can I get a link?

https://www.mql5.com/ru/forum/86386/page3570#comment_53975342

[Forester]

Maxim Dmitrievsky #:
They can improve a lot if you do it in a meaningful way, like in Alexei's article.

I read it yesterday. I haven't tested price speeds and accelerations, volatility yet. It is unlikely to find something else without connection to the exchange, especially for forex instruments. Besides, in the article trading on D1, it seems. It's a bit rare.
And it's easy to make recession marks on history. In real life we will notice it only by news (with a sharp collapse), or in 2-3 months by charts, when we realise that the correction does not end in any way.

[Forester]

Maxim Dmitrievsky #:
h ttps:// www.mql5.com/ru/forum/86386/page3570#comment_53975342

Interesting. They wrote there that in about half of the models these new features gave improvements. But they didn't tell about the other half.
If in the other half it became worse, then on average probably nothing has changed. Perhaps these improvements and deteriorations are related to randomiser, as it happens at different restarts of training on the same data - it turns out to be better or worse.
But in principle the chips are useful, probably.

[Maxim Dmitrievsky]

Forester #:
Interesting. They wrote there that in about half of the models these new features gave improvements. But they didn't mention the other half.
If the other half got worse, then on average probably nothing has changed. Perhaps these improvements and deteriorations are related to randomiser, as it happens at different restarts of training on the same data - it turns out to be better or worse.
But in principle the chips are useful, I guess.

I didn't read too much, but, logically, different modes with the same features will be worse/better. I observe the same thing: half of the models work well, half not so well. Different patterns.

So I additionally introduced different types of markups, like momentum strategy and return to the mean. And overlaid these markups on different modes. The best patterns so far have come out with the reversion-to-average strategy. Momentum strategies seem to be more capricious. The former work better on flat instruments/modes, the latter on trending ones.

[Forester]

Maxim Dmitrievsky #:
I didn't read too much, but, logically, different modes with the same features will be worse/better. I observe the same thing: half of the models work well, half not so well. Different patterns.

So I additionally introduced different types of markups, like momentum strategy and return to the mean. And overlaid these markups on different modes. The best patterns so far have come out with the reversion-to-average strategy. Momentum strategies seem to be more capricious. The former work better on flat instruments/modes, the latter on trending ones.

So we need to train the model to identify the mode:
1) flat - counter-trend trade, even without MO.

2) trend up, we trade on the trend, we can do it without MO as well

3) trend down

[Maxim Dmitrievsky]

Forester #:

So we need to train the model to determine the mode:
1) flat - countertrend trade, even without MO.

2) trend up, we trade according to the trend, we can also trade without MO

3) trend down

It is impossible to understand at once what will work, so the selection of modes and markups is relevant. Clustering criterion/number of clusters is also important. All this stuff gives some results only in aggregate. You can flexibly tune through MO, but you will get bored in the optimiser. Especially if your own tester is 100 times faster than in the platform. You don't even need a tester at the selection stage, you can compare errors, which is fast. If you come across a good result, you save the parameters and recheck it.

Theoretically, and nothing prevents in practice, an external macroeconomic indicator can be taken to determine the regime, not necessarily derivatives of prices/moments of price distribution. And the main model will be based on prices for trade.

[mytarmailS]

I don't know if it's okay to post this here, if it is, delete it or ban it or whatever you have to do...

There is such a contest https://bitgrit.net/competition/22 It is not a kaggle, but there are good prizes, very small date set, 63 observations, time series data.

Try anyone who feels confident...

[Grigori.S.B]

mytarmailS #:

I don't know if it's okay to post this here, if anything delete it or ban it or whatever you have to do....

There is such a contest https://bitgrit.net/competition/22 It is not a kaggle, but there are good prizes, very small date set, 63 observations, time series data.

Try anyone who feels confident...

It's for astronauts.
))))

[mytarmailS]

Grigori.S.B #:

It's for astronauts.
))))

AND WHO ?????

what does that change and where?

[Renat Akhtyamov]

mytarmailS #:

I don't know if it's okay to post this here, if anything delete it or ban it or whatever you have to do....

There is such a contest https://bitgrit.net/competition/22 It is not a kaggle, but there are good prizes, very small date set, 63 observations, time series data.

Try anyone who feels confident...

Artificial intelligence again.

I can do such a feature in one or two times, but I can't help the states.)

I don't understand what to look for?

The simplest device...

For example, yesterday I made a device to measure and visualise the magnetic field pattern.