viability of predicting multiple tokens at the same time

[Farbdose]

I was looking for this kind of research and just found your repo and you actively working on it.

Did you consider training the LSTM model on all tokens that are offered on a specific crypto exchange at once? I'm curios if this would have any effect on the accuracy as one could assume that the money mostly stays in the exchange meaning the tokens influence each other.

The endgoal would be to not predict if a token goes up or done but to predict which token will increase the most in the selected timeframe.

The trading process could be one trade per day and it switches to the token the model predicts will have the highest increase in closing price - for simplicity the model could be trained on the same time period one plans to trade on - like training on daily time frame, the model would have to identify which of the used coins will most likely have the highest increase in the next bar/day.

As you are currently doing research in this field could I ask you to give me your opinion on this approach?

[shahgem]

Firstly, I appreciate your interest in my study work.

Regarding the point you made about training the LSTM model on all tokens offered on a specific crypto exchange at once, I did not initially train the model on all tokens at once but same parameters were used on all tokens. However, after considering your suggestion, I conducted further experiments and trained the LSTM model on all tokens at once using the same parameters.

Interestingly, the results did not show a significant difference compared to training the model on individual tokens separately. This suggests that, for the specific exchange and dataset used in this study, the influence of tokens on each other may not have a significant impact on the accuracy of the LSTM model.

Nonetheless, I believe it is an interesting and valid point to consider particularly with larger datasets and different exchanges. It is possible that in other scenarios, the tokens could influence each other's values, and training the LSTM model on all tokens simultaneously could lead to better predictions.

In summary, while training an LSTM model on all tokens offered on a specific crypto exchange at once is a valid approach, its effectiveness may vary depending on various factors. Therefore, it would be worth exploring in future research to determine its potential benefits and limitations.

[Farbdose]

Thank you for your inside. I will investigate this further on my own and cite your repository as source for the code. For future reference this was my own attempt at training with multiple tokens: https://colab.research.google.com/drive/18aZx5G0Jq7nmxCR5pCNbIaNf56-FityZ?usp=sharing

[shahgem]

The optimal approach to investigating each model depends on the specific datasets used. Instead, it is important to carefully consider the limitations of each model and strive for clarity in order to meet the desired goals.

I am pleased to hear that my study has been beneficial to you in some manner.