Predicting bitcoin trends reversals with machine learning methods

In recent years, Bitcoin has become the most popular digital asset in the cryptocurrency market. Its prices are highly volatile due to rapidly increasing investor interest, making it difficult to predict price movements. In this context, predicting the direction of Bitcoin prices can assist investors in determining their investment strategies. Machine learning techniques are widely used in predicting price directions and making trading decisions in financial markets. Machine learning models can be developed using time-dependent data and financial indicators to forecast future price movements. Bitcoin price movements can be influenced by numerous factors in the cryptocurrency market. Therefore, determining the accuracy and success of machine learning techniques is crucial. The aim of this study is to predict trend reversals in Bitcoin price movements by using machine learning techniques and compare the success rates of these techniques. For this purpose, the study focuses on points where the trend changes. A three-class prediction study is conducted in order to classify the price direction as 'buy', 'sell', or 'hold'. Daily data collected from 'binance.com' between September 1, 2017, and April 1, 2023, including the open, close, high, low prices and Bitcoin trading volume are used to identify trends and calculate financial indicators. After the data examination stage, dimension reduction is performed by using principal component analysis to avoid multicollinearity issues of independent variables. The 'buy', 'sell', and 'hold' classes are balanced through under-sampling. Extreme Gradient Boosting (XGB), Random Forest (RF), Random Tree (RT), Artificial Neural Networks (ANN), Support Vector Machines (SVM), Bayesian Networks (BN), and K-Nearest Neighbors (KNN) models are developed. The results are evaluated by using precision, recall, specificity, F1 score and accuracy metrics. Since ANN, SVM, BN, and KNN models show significantly lower performance compared to tree-based algorithms, the comparison in the study focuses on the results of XGB, RF, and RT models. The study concludes that the XGB model exhibits higher success compared to other models. Keywords: Bitcoin trend prediction, cryptocurrency price analysis, machine learning, classification