Machine learning in trading: theory, models, practice and algo-trading - page 3347

New comment
 
fxsaber #:

"Hello World!" in the area of understanding raw data - to write a script that will show the maximum possible profit on a historical interval.

If you don't have it, then it is unclear what you are doing.

Not only everyone can reach this point, studying the MOE. It is not written anywhere. Usually it is very simple desires to cram more signs and more layers into the NS.
 
nevar #:

NHITS and the lightGBM also has lower RMS then the TimeGPT in daily and hourly data. https://valeman.medium.com/what-truly-works-in-time-series-forecasting-the-results-from-nixtlas-mega-study-78eda5133622



Have you tried Conformal Prediction ?

https://valeman.medium.com/how-to-predict-full-probability-distribution-using-machine-learning-conformal-predictive-f8f4d805e420

https://github.com/valeman/awesome-conformal-prediction#papers-time-series

Hello, not yet. Thank you for such an amazing book and articles, I will definitely read it.
 
Maxim Dmitrievsky #:


So it turns out that the pattern itself is at the level of the spread, or how to interpret it? That is, it does not cover trading costs.

So it is, on mashki achieved a uniform loss on each trade equal to the spread)))))) At zero spread even profit))

 

Maxim Dmitrievsky#:


So it turns out that the pattern itself is at the level of spread, or how to interpret it? That is, it does not cover trading costs.

The spread has nothing to do with it.

We just take the statistics for H1

> quantile(abs(na.omit(diff(CLOSE))), probs = seq(0, 1, 0.01)) 
       0%        1%        2%        3%        4%        5%        6%        7%        8%        9%       10%       11%       12% 
0.0000000 0.0000100 0.0000200 0.0000300 0.0000400 0.0000500 0.0000600 0.0000700 0.0000800 0.0000800 0.0000900 0.0001000 0.0001100 
      13%       14%       15%       16%       17%       18%       19%       20%       21%       22%       23%       24%       25% 
0.0001200 0.0001300 0.0001400 0.0001500 0.0001700 0.0001800 0.0001900 0.0002000 0.0002100 0.0002200 0.0002300 0.0002400 0.0002500 
      26%       27%       28%       29%       30%       31%       32%       33%       34%       35%       36%       37%       38% 
0.0002600 0.0002700 0.0002800 0.0002900 0.0003000 0.0003100 0.0003200 0.0003400 0.0003500 0.0003700 0.0003800 0.0003900 0.0004000 
      39%       40%       41%       42%       43%       44%       45%       46%       47%       48%       49%       50%       51% 
0.0004200 0.0004300 0.0004400 0.0004500 0.0004700 0.0004800 0.0004900 0.0005100 0.0005300 0.0005400 0.0005500 0.0005700 0.0005900 
      52%       53%       54%       55%       56%       57%       58%       59%       60%       61%       62%       63%       64% 
0.0006000 0.0006200 0.0006400 0.0006600 0.0006800 0.0006900 0.0007100 0.0007300 0.0007500 0.0007700 0.0007900 0.0008100 0.0008300 
      65%       66%       67%       68%       69%       70%       71%       72%       73%       74%       75%       76%       77% 
0.0008600 0.0008800 0.0009100 0.0009300 0.0009600 0.0009800 0.0010100 0.0010300 0.0010600 0.0010900 0.0011300 0.0011700 0.0012100 
      78%       79%       80%       81%       82%       83%       84%       85%       86%       87%       88%       89%       90% 
0.0012500 0.0012900 0.0013300 0.0013700 0.0014200 0.0014600 0.0015100 0.0015730 0.0016300 0.0017000 0.0017700 0.0018500 0.0019300 
      91%       92%       93%       94%       95%       96%       97%       98%       99%      100% 
0.0020200 0.0021500 0.0022900 0.0024400 0.0026400 0.0029300 0.0032600 0.0037200 0.0048404 0.0173300

and stupidly see at what value of price increment your "profitable" forecasts become unprofitable, i.e. at 10 pips of 4 digits only 25% of market movements become potentially profitable. This is with an error-free forecast!

 
СанСаныч Фоменко #:

Spread has nothing to do with it.

We just take statistics for H1

and stupidly see at what value of price increment your "profitable" forecasts turn into unprofitable ones, i.e. at 10 pips of 4 digits only 25% of market movements become potentially profitable. This is with an error-free forecast!

You do not understand what I am writing about

When marking with the spread, 0% of trades are unprofitable. And it does not matter whether it is calculated by average price + spread, or by Saber ticks on bid and ask separately. On average, the result is comparable.

you can calculate by ticks later, if you are a fierce scalper and work in 1-2 dts, I do not particularly like such TSs

 

Draw a diagram-distribution of deals, where on the horizontal line is the profit of closed positions, on the vertical line is the number of closed positions.

For narrow spread and wide spread.

 
Maxim Dmitrievsky #:

You don't understand what I'm writing about

When marking with the spread, 0% of trades are unprofitable. And it does not matter whether it is calculated by average price + spread, or by Saber ticks on bid and ask separately. On average, the result is comparable.

you can calculate by ticks later, if you are a fierce scalper and work in 1-2 dts, I do not particularly like such TSs

My markup is price increment.

Take your markup and look at quantile What profit your markup is designed for? Compare it with the statistics.

 
СанСаныч Фоменко #:

My markup is the price increment.

On the same list.

Forum on trading, automated trading systems and testing trading strategies

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

fxsaber, 2023.12.10 17:57

Talking about spread, timeframes and Japanese candlesticks is about the same thing.

 
СанСаныч Фоменко #:

My markup is the price increment.

Take your markup and look at the quantile. How much profit is your markup designed to make? Compare it to the statistics.

No, there's no problem with that. It doesn't matter what the profit margin is. What matters is the classification error. It grows when spread is added to the training or remains the same.

But the model does not start working better when the spread is taken into account in the markup, it does not give a profit, but without the spread it works the same way as if it was trained without it. That is why I put the spread, conditionally, to the classification error. That is, the response of the model does not allow you to beat it.

Taking the spread into account in the markup means the length of trades that exceed it. That is, I make the trades longer, then train them, and the result of testing on the increased spread is almost the same as the result of another model trained on shorter trades.

It turns out to be a rather unambiguous conclusion that on my signs, let's say, MO cannot beat the spread.

But sometimes it can, with certain machinations related to kozul. That is, if there is some stat. indicator of deduced "reliability" of signals, then they work also when the spread increases.

 
Maxim Dmitrievsky #:

Nah, there's no problem with that. It doesn't matter what the profit margin is. What matters is the classification error. It grows when adding spread to training or remains the same.

But the model does not start working better when the spread is taken into account in the markup, it does not give a profit, but without the spread it works the same way as if it was trained without it. That's why I put the spread, conditionally, to the classification error. That is, the response of the model does not allow you to beat it.

Taking the spread into account in the markup means the length of trades that exceed it. That is, I make trades longer, then train them, and the result of testing on the increased spread is almost the same as the result of another model trained on shorter trades.

It turns out to be a rather unambiguous conclusion that on my signs, let's say, MO cannot beat the spread.

But sometimes it can, with certain machinations related to kozul. That is, if there is some stat. indicator of deduced "reliability" of signals, then they work also when the spread increases.

Itdoesn't matter what profit is calculated for.What matters is theclassification error.

Because of this approach you "correctly" classify potentially losing trades. In reality, the situation is much worse not only because of the spread. In a real EA to achieve from "correct" classification to a profitable system remains a problem, as it is not surprising.

1...334033413342334333443345334633473348334933503351335233533354...3627
New comment