Python Forecasting Bitcoin (BTC) with Machine Learning

[BRN]

Hi everyone,
I apologize to the mod if I posted in the wrong section.

For my exam of Machine Learning, I would like to implement a part of the work presented in this paper. In this work, the authors used two ML methods in cascade for forecasting Bitcoin. Starting from the initial data, they predicted the five main BTC indicators via SVR (Support Vector Regression) and the latter were then the inputs used to predict the price of Bitcoin via LSTMNN (Long Short Term Memory Neural Networks).

There are two things that I don't understand:

• Starting with a dataset of about 2600 rows, 80% of them are used for Trainig and only 20% for the prediction coming out of SVR. This data are ulteriorly separated by incoming in LSTMNN and follows that the prediction in this stage is made on a truly reduced sample compared to the starting one. It's not a problem?
• If the input to LSTMNN are only the indicator predicted by SVR, how is it possible to forecasting the price of the BTCs? At this stage there is only the X matrix of the indicators, but there is not price vector...

Can anyone clarify my ideas?

Thanks so much!

 

[Superintendent]

Create an artificial market have 30 to 40 simulated miners and sellers, have that market resemble how BTC has been doing, teach the Algorithm Off of that, introduce it to the real BTC market and continue to train it once the First Bit of training with the simulated market has been finished,

 

[BRN]

Create an artificial market have 30 to 40 simulated miners and sellers, have that market resemble how BTC has been doing, teach the Algorithm Off of that, introduce it to the real BTC market and continue to train it once the First Bit of training with the simulated market has been finished,

Thanks for the reply.

Ok, I tried to implement it. Through SVR model, I predict the technical indicators which, once inserted in LSTM model, providing the price of the BTCs. Everything works, but the results are really bad ...

The code is too long to post it here, but here there are my Jupyter notebook and the dataset file.

If someone could give me some advice on how to improve it / correct it, I would be really grateful.

Thanks so much!

 

[Superintendent]

Thanks for the reply.

Ok, I tried to implement it. Through SVR model, I predict the technical indicators which, once inserted in LSTM model, providing the price of the BTCs. Everything works, but the results are really bad ...

The code is too long to post it here, but here there are my Jupyter notebook and the dataset file.

If someone could give me some advice on how to improve it / correct it, I would be really grateful.

Thanks so much!

Sorry, i am not that good in the aspects of Machine learning, just thought of that during school and i thought i should share the information to you

 

[Nik_2213]

Beware: 'Real Life' may totally trump logical predictions.
IIRC, Bitcoin value tumbled this morning as eg Russian miners 'cashed in', perhaps fearing financial sanctions following trouble along Ukraine's borders...

May be a 'Pump & Dump' ploy, may simply be 'Duck & Cover'...

Like the many TV ads for 'Gold Bullion', too many bit-coin mining proponents fail to mention that value of investment may go down as well as up. Worse, if playing with 'futures', your investment may be wiped out, and leave you with significant debts...

Due Care, Please ??

 

[BRN]

I don't have to invest money in BTCs. As I have already specified in my first post, I just have to present a project for an exam.

The predicted price values are compared with the real values present in the test dataset. By calculating the MAE (Mean Absolute Error), I get a very high value. I would like to understand if the problem is how I set the model or for some other reason.

I ask to anyone who is familiar with Machine Learning and Big Data to examine my Jupyter Notebook and tell me what he thinks.

 

[Nik_2213]

Ah, you assume a 'Mature Bourse'.
Go for it...

 

[Tom.G]

The predicted price values are compared with the real values present in the test dataset. By calculating the MAE (Mean Absolute Error), I get a very high value.

Ahh! I see you are rapidly learning the vagaries of Bitcoin.

It is a bit like the Stock Market combined with a late-nite Infomercial^*, but without the U.S. Securities and Exchange Commission to keep an eye on it.

(At least you are learning Economics without going broke. )

I rather liked @Nik_2213 's reply above:

Beware: 'Real Life' may totally trump logical predictions.

* in·fo·mer·cial
/ˈinfōˌmərSH(ə)l/

noun
a television program that promotes a product in an informative and supposedly objective way.

 

[valenumr]

I don't have to invest money in BTCs. As I have already specified in my first post, I just have to present a project for an exam.

The predicted price values are compared with the real values present in the test dataset. By calculating the MAE (Mean Absolute Error), I get a very high value. I would like to understand if the problem is how I set the model or for some other reason.

I ask to anyone who is familiar with Machine Learning and Big Data to examine my Jupyter Notebook and tell me what he thinks.

Humans are irrational. Especially in crypto these days. I would be highly impressed by an AI that could accurately predict crypto price trends.

 

[russ_watters]

Humans are irrational. Especially in crypto these days. I would be highly impressed by an AI that could accurately predict crypto price trends.

Yeah, AI that could accurately predict price trends in pretty much anything would make its creator exceptionally wealthy.

 

[BRN]

Maybe I explained myself wrong.

I don't want an AI who is able to predict the price of the BTCs accurately. I don't have to invest in Crypto. This is just a project to be presented for my ML exam. I chose this topic because I don't want to present the usual project that generally most of the students of the course present (image recognition, sub-particles analysis, etc.). I want to present something different.

I started from this article where the authors say that the two models in cascade SVR+LSTM return a better result than the single one LSTM. Well, from how I implemented it, this result doesn't happen. For this reason I ask someone to look at my code and give me an opinion of it.

• Is my code correct?
• Is it correct to put SVR and LSTM in cascade in this way?
• Is it right to make LSTM training with all the data available?
• Did I overfitting?
• Does not implement a Montecarlo system to estimate LSTM hyperparameters is relevant?
• and so on...

All these questions are technical questions and no matter whether they are related to a model that manipulates crypto data, sub-particles data, potatoes or other.

If you don't believe that I should take an exam, I can't do anything. Where I live, in the universities the exams are made to obtain a degree.

I didn't think I had to clash even against prejudices...

 

[Office_Shredder]

Just a quick starting question - is your code attempting to compute the exact same technical predictors, fit the exact same model, and using the exact same training and test data?

 

[willem2]

The predicted price values are compared with the real values present in the test dataset. By calculating the MAE (Mean Absolute Error), I get a very high value

I think you should only look at relative errors for bitcoins, or share or commmodity prices. A 1% error should be weighted the same when bitcoin is at 1$ as when it is at 10000$

 

[BRN]

Just a quick starting question - is your code attempting to compute the exact same technical predictors, fit the exact same model, and using the exact same training and test data?

The technical predictors are the same used by the authors of the article. I also tried to add others: MACDS, MACDH and ROI, but the result does not change.
My dataset is more updated, but I tried to use the data of the same period indicated in the article and I always get that LSTM works better than SVR + LSTM.
No I have idea of how the authors have implemented the SVR+LSTM model, because on this stage, they have not been very detailed in the description. For this reason I am asking your opinions here. I just know that my code is OK. The only difference I know is that I have not implemented a Montecarlo method to estimate LSTM hyperparameters.

I think you should only look at relative errors for bitcoins, or share or commmodity prices. A 1% error should be weighted the same when bitcoin is at 1$ as when it is at 10000$

I use the Mean Absolute Error because in the article it is used, but I can also use the Relative one. I don't think it's a problem.