Cryptocurrency Trend Prediction Based On Imbalanced Data

As a new investment product,cryptocurrency is sought after in the market due to its good technology application prospects and high return on investment.Like other investment products,the market desperately needs an automated analysis tool to analyze the historical data of cryptocurrencies and analyze the future of cryptocurrencies.Machine learning methods are widely used in various fields,and the cryptocurrency market is no exception.Among them,ensemble learning methods have advantages such as high efficiency,excellent performance,and strong robustness compared to traditional machine learning methods.More and more researchers are beginning to pay attention How to better apply integrated learning to the cryptocurrency market.In the cryptocurrency data set,most of the samples belong to the fluctuation interval,and only a few samples belong to the rising or falling,and the rising and falling samples are the important research objects of the cryptocurrency trend prediction.The target of general machine learning algorithms is to maximize the model accuracy.In data imbalanced data sets,model learning will be biased towards the majority of samples,resulting in a reduction in the classification performance of the minority of samples.When using machine learning methods to analyze cryptocurrencies,solving data imbalances is an important part that cannot be ignored.The purpose of this paper is to improve the classification performance of ensemble learning models in the cryptocurrency market.Based on this target,we first proposed SC-SMOTE over-sampling algorithm.SC-SMOTE algorithm based on clustering and feature distribution to solve the data imbalance from the data level,and then combined the SC-SMOTE algorithm and weight balance factor on the Boosting ensemble learning.A balanced ensemble learning algorithm SCSMOTE-Boost algorithm for the cryptocurrency market is proposed.The algorithm clusters the data set selects representative seed samples,and performs over-sampling based on the distribution of the characteristic data in the clusters,increasing the emphasis on the spatial distribution of the samples and the distribution of the characteristic data;at the same time,a The sample weight balance factor improves the importance of minority classes in the ensemble learning process.SCSMOTE-Boost algorithm is optimized from three aspects:data level,ensemble learning algorithm and weight adjustment.Experiments on multiple types of cryptocurrencies have proved the effectiveness and stability of the SCSMOTE-Boost algorithm in cryptocurrency trend prediction situations.