Which algorithm is being used for the signals on Telegram?

[make-j64]

Could you tell me what algorithm the configuration you're running at https://t.me/intelligent_trading_signals is using? I've trained using config-sample-v0.5.0, but the results are vastly different from the signals in https://t.me/intelligent_trading_signals. Thank you for the excellent project.

[asavinov]

The results depend not only on the algorithms but also on features defined. The Telegram bot has used different configurations for experimenting and test purposes. Currently it uses neural network and linear classifier. Previously it used gradient boosting. The algorithm parameters are defined in model_store.py in models dict.

[make-j64]

Can you share the config of Telegram bot?

[adiif1]

+1

[iploshchik]

would also appreciate to get the config used shared, could be a good starting point for new comers to check against provided signals in telegram

[tom634]

The results depend not only on the algorithms but also on features defined. The Telegram bot has used different configurations for experimenting and test purposes. Currently it uses neural network and linear classifier. Previously it used gradient boosting. The algorithm parameters are defined in model_store.py in models dict.

Do you still use neural network and linear classifier or you changed it to something different?

[asavinov]

Currently only NN

[tom634]

Why is that? NN is in your opinion the best?

[asavinov]

I did not see big benefits from using linear classifier but it is slow to train. Actually I am using NN with only slightly more parameters than LC. For 1 hour data, I use SVM just because it is possible to train it in reasonable time. I also generate a lot of input features and in this case it is much less important which algorithm to use. It is more important to find a good trade strategy based on this (and other if available) indicators, that is, to define conditions for buying and selling. For example, how to trade in case there are intelligent indicators (scores) for 1 minute, 5 minute and 1 hour data.

[issacnumb]

feature is important

[issacnumb]

In fact, I think what the author wants to express is that you need to consider algorithm indicators in different time frames at the same time, such as issuing long signals at the same time in 5 minutes, 15 minutes, and 1 hour, and then decide to buy.