Machine learning in trading: theory, models, practice and algo-trading - page 1602

[...]

Hi, advise some libraries or ML frameworks that work well with classification between, say, 10 different classes, ideally in C#

Running ML.NET on his input, he identified such methods as the most accurate

- LGBM classifier.

- Maximum Entropy classifier.

But ML.NET has some limitations, so he would like to try other options.

This is not related to the market right now, but will probably be used in the future.

Example data.

Label   Pitch   Energy  RMS     ZCR     Centroid        Spread  Flatness        Noiseness       RollOf  Crest   Entropy Decrease        C1      C2      C3      C4      C5      C6      M1      M2      M3      M4      M5      M6      M7      M8      M9      M10     M11     M12     M13

08      195.91840       749479.40000    663.49990       0.06797 5960.71800      5623.56900      0.45596 0.48241 11892.42000     113.03180       0.81972 -0.01187        0.60059 0.99703 1.12502 1.38532 1.41049 1.34596 174.31960       5.43771 0.53368 1.08706 1.28929 -0.27730        0.22525 -0.32192        -1.06489        0.18286 0.12653 -0.36697        0.04997

07      89.38548        264804.20000    427.55540       0.06677 5575.63400      5633.31900      0.41806 0.46413 11484.85000     167.75290       0.79212 -0.17176        0.58641 1.03448 1.14216 1.40824 1.38114 1.41114 174.84360       4.86947 0.42199 1.17480 1.67603 -0.33066        0.54447 -0.31041        -0.75327        -0.04792        0.82607 0.23418 0.16688

[Aleksey Vyazmikin]

...:

Hi, advise some libraries or ML frameworks that work well with classification between, say, 10 different classes, ideally in C#

Running ML.NET on his input, he identified such methods as the most accurate

- LGBM classifier.

- Maximum Entropy classifier.

But ML.NET has some limitations, so he would like to try other options.

This is not related to the market right now, but will probably be used in the future.

Example data.

Have you triedCatBoost?

[Andrei01]

Maxim Dmitrievsky:
The fortuneteller did not see the chelists, i.e. anonymously. Just asked her short questions by different people and they evaluated the answer. No one knows each other. I just wondered what the psychological effect was. Thought I would fall out nonsense on the cards, and they're on point.)

The same place the amoeba "knows" from. What's the fortune teller?)

[Evgeny Dyuka]

I did manage to train a neural network. For BTCUSD it predicts movement every minute for the next 15 min (without reference to candlesticks). The neural response is binary up/down and expressed as a numerical value in the range from -70 to +70, this is not a price prediction, it is a degree of confidence in the movement.
Now in the real market, the result is higher than expected. During the backtest 68% of successful answers turned out to be much better. I trained without crutches and hints, i.e. no external influence, everything I trained worked.
The data has been prepared with MQL5 bot, TensorFlow + Keras neuron, now it sends my predictions to Telegram channel. I don't give the link here, it seems to be impossible, but if I can let me know.

In fact I got an indicator which gives out a value on every minute candle. At values of 30 and above I can try to trade)

I will answer the questions, but I will leave know-how about data preparation for training ...

[Andrei01]

Evgeny Dyuka:

On the backtest gave 68% of successful answers and life was much better.

That says nothing, as you know. What is the expectation of winning?

[Evgeny Dyuka]

Andrei:

That says nothing, as you know. What is the expected payoff?

Of course it doesn't say anything, but the real market does. The forecast is absolutely correct, the guessing of trends and forthcoming reversals is very good, so to say. better than all known indicators. All this is public.
Mat expectation was not counted.

[Andrei01]

Evgeny Dyuka:
But the real market says, the forecast is absolutely adequate, the guessing of trends and upcoming reversals is very good, so to say... Better than all the known indicators. All this is public.

The percentage of guessing does not say anything, there can be 99% of profitable guesses and 1% of unprofitable, which will cover all the profits.

Evgeny Dyuka:
Mat expectation was not counted.

Why so?

[Renat Fatkhullin]

Has anyone used the built-in python scripts and Metatrader5 library for Python yet?

• Documentation: MetaTrader for Python
• New version of MetaTrader 5 build 2340: Managing account settings in the tester and expanding integration with Python
• MetaTrader 5 Python User Group - how to use Python in Metatrader

[Evgeny Dyuka]

Andrei:

Percentage of guesses does not say anything, can be 99% of guesses are profitable and 1% of unprofitable, which will cover all the profits.

Why so?

You can convince me that I'm wrong. This project I did six months for 16 hours a day and got the result. I can not show you because you are waiting for the link that I give will be advertising casinos.

A counter question, and where else to discuss this topic as on the profile forum?
Maybe for safety we equip the delegate from the forum, he put on a protective suit, shodt on link, observe and tell - there or not a casino.
After that we'll discuss.

[Evgeny Dyuka]

Andrei:

Percentage of guesses does not say anything, can be 99% of guesses are profitable and 1% of unprofitable, which will cover all the profits.

Why?

The percentage of guesses is not important in this topic, there are other categories. The main thing here is whether the network is learning or not. If it shows signs of learnability, the door is open, and then the quality of the prediction can be increased infinitely. Everything will depend on equipment and time.