Cryptocurrency price forecast
I.Our purpose is to forecast
1. bitcoin close price in the next day (short term forecast);
2.the trend for the relative long term (the next week or month);
3. Compare different models and ascertain the accuracy of the direction of Bitcoin in USD can be predicted.
II. Features already in the dataset
| btc_market_price | btc_total_bitcoins | btc_market_cap | btc_trade_volume | btc_blocks_size |
| btc_avg_block_size | btc_n_orphaned_blocks | btc_n_transactions_per_block | btc_median_confirmation_time | btc_hash_rate |
| btc_difficulty | btc_miners_revenue | btc_transaction_fees | btc_cost_per_transaction_percent | btc_cost_per_transaction |
| btc_n_unique_addresses | btc_n_transactions | btc_n_transactions_total | btc_n_transactions_excluding_popular | btc_n_transactions_excluding_chains_longer_than_100 |
| btc_output_volume | btc_estimated_transaction_volume | btc_estimated_transaction_volume_usd |
III. Features plan to be included
1.Sentiment is one of determinant of bitcoin price. To pull the data like Twitter sentiment ratio, and Wikipedia search queries (which shows the degree of public interest in Bitcoins). Hash rate from blockchain measures the mining difficulty (Already got).
2.USD and Euro exchange rate, as Euro represents the general level of prices. Brent Crude oil price, gold price, S&P 500 which implies the general state of the economy and also are substitute for Bitcoin as investment to get profits.
IV. Steps
1. Download the predictors that capture sentiment and economy condition. And combined it with our “Bitcoindatset.csv” dataset.
2.Clean the dataset remove the series that have 0 standard error, and do missing value imputation; etc.
(2. Feature selection (subset selection, Lasso), using cross-validation to select the subset that have the smallest prediction error.we can use statistical method to do this.)
3. Machine learning method
1)Classification method (Gaussian NB, KNN, Logistic regression) to predict the direction and report f1 score.
I read paper that people use Long Short Term Memory network to predict the direction of bitcoin and achieve the highest accuracy.
2)Regression method to predict the price of next day close price. (ARIMA, GARCH for short term, Bayesian Vector autocorrelation for long term)
Notice that:
As the chart shows, before 2013, Bitcoin price was stable, but still have variations, we can try to fit the model to the whole dataset, and also select different intervals to fit the model, treating different time period separately. Sample itself is an variable for our model.
