Long Short-Term Memory for Bitcoin Price Prediction

Abstract

With time-series data being prevalent everywhere, there is a need to predict this data accurately. This kind of data includes weather data, financial data such as stock price, and cryptocurrency price. Most of the trades in the stock market in this day and age are being made using artificial intelligence. An estimated 50% of trades were done using an algorithm, which increased to 60% in 2020 [1]. This highlights the demand for reliable and accurate predictions. The prediction of the price is very challenging. Some success has been seen when predicting stock prices, but not many studies have been done on cryptocurrency. Cryptocurrency, specifically Bitcoin, has seen a substantial increase in popularity, and the price has reflected this popularity. The price also follows patterns specifically when reaching new all-time highs. In this work, an Artificial intelligence is created and trained on the previous data to observe these patterns and predict the next price. The artificial intelligence chosen for this subject is Long short-term memory (LSTM). LSTMs are capable of finding patterns in time series data. LSTM solves the vanishing gradient problem present in the RNN (Recurrent Neural Network). The Market Price of Bitcoin is used as input here. The data values for input range from 20,000 up to 65,000 in testing. Once an optimal starting point is found, there is an 80/20 split of data, 80 percent of the data is used for training and 20 is used for testing. With the data being split, one of the most important jobs is figuring out the optimal lags (how far back into the past) when used to predict values. This range for this experiment is set to ten previous price days. Epochs (number of iterations) and Batch size (how much of the training data is used per epoch) are tested at different values to find optimal solutions. With batch size values such that batchSize ∈ {20, 21…26} and epochs such that epochs ∈ {10, 20….70}. Overfitting is hard to detect and thus can be an issue with too many epochs and smaller batch sizes (smaller means more of the training data is used). Too little and the LSTM will not learn the data patterns and thus will not have good accuracy. This is why different configurations are used in the experiment to maximize accuracy. This LSTM was used to achieve a Mean Absolute Percentage Error score of 3.23% and a Root Mean Squared Error score of 1892.87 when predicting next-day prices throughout 350.