CS460 | Group 3 | Cryptocurrency prediction

Introduction

The proposal and midway report are on the following link.

This is the final report for the group 3's project for CS460. The goal of the project was to predict the values of crypto currencies. We have tried out various sources of data, which could be affecting the data like interest from Google trends, and values of other crypto currencies, and have come up with considerable predictions and models.

The input data includes the value of all the crypto currencies for the day before the day which is being predicted for, and also google trends interest value for the earlier day.

Models and Predictions

From initial experimentation (discussed in midway) we found that the value of the cryptocurrency did not change by a huge value over the period of a day, so if the input contains the data of the last day. A linear model sets the coefficient of other inputs to ~0 and the last day value of the coin to the same value. Which while is technically still correct is not really providing any new information. Our first approach was to predict the difference of the value instead the value of the coin, but the models seem to not predict that with very good accuracy. Second and the successful approach was to not provide the last days value in the feature vector and asking the model to predict the value.

None of the models in the rest of the report, provide the value of the coin which is being predicted into the feature vector. With some exception we are getting very impressive results, if we would say so ourselves.

Time vs Bitcoin value — Bitcoin value with time

Linear model predictions of Bitcoin value — Linear Model predictions

Neural network predictions of Bitcoin value — Neural Network predictions

Link to Neural Network model

Time vs Ethereum value — Ethereum value with time

Linear model predictions of Ethereum value — Linear Model predictions

Neural network predictions of Ethereum value — Neural Network predictions

Link to Neural Network model

Time vs Binance value — Binance value with time

Linear model predictions of Binance value — Linear Model predictions

Neural network predictions of Binance value — Neural Network predictions

Link to Neural Network model

Time vs Cardano value — Cardano value with time

Linear model predictions of Cardano value — Linear Model predictions

Neural network predictions of Cardano value — Neural Network predictions

Link to Neural Network model

There wasn't much variance in the value for the model to predict anything properly.

Time vs Tether value — Tether value with time

Linear model predictions of XRP value — Linear Model predictions

Predictions from neural network were constant, we are still investigating why that is.

All the deviations in the price were in the very late stages as can be seen here. While the interest values and other parameters varied in the initial phases, so the models mostly predicts lower values.

Time vs Solana value — Solana value with time

Predictions were really bad, we are still investigating why that is.

Time vs Polkadot value — Polkadot value with time

There wasn't much variance in the value for the model to predict anything properly.

Time vs USDCoin value — USD_Coin value with time

Time vs Dogecoin value — Dogecoin value with time

Linear model predictions of Dogecoin value — Linear Model predictions

Predictions from neural network were constant, we are still investigating why that is.

Conclusion

While we initially considered to extract information from tweets, we were not able to collect enough tweets to run any analysis, primarily because of the sheer amount of tweets which we are interested in.

The values of cryptocurrencies can be used to predict the behaviour of a particular cryptocurrency. The exact relations can be found by analysing coefficients of Linear models.
We have linear models to predict value of the next day based on parameters which can be collected on an earlier day.

Appendix

Github link to code