Research On Cryptocurrency’s Returns,Risk Spillover Effects,and Portfolios Based On Machine Learning

As a key component of digital assets,cryptocurrency has brought changes to the traditional financial system and become a unique and important new financial asset.The decentralized nature and underlying blockchain technology give cryptocurrencies the potential to hedge against some economic uncertainty and inflation,and also provide investors with diversified financial assets.Cryptocurrency’s returns exhibit remarkable highs compared to traditional financial assets,coupled with elevated volatility and non-normal distributions.The existing studies on cryptocurrency’s price influencing factors,prediction methods,risk spillover effects,and portfolio strategies are inadequate.Therefore,this research provides a more in-depth investigation and analysis of the returns and risks of cryptocurrency.This research addresses the key scientific issues in the field of cryptocurrency from the following three aspects.In terms of cryptocurrency’s returns,how to mine the pricing characteristics of cryptocurrency and measure its price influencing factors from a multi-dimensional perspective? In terms of cryptocurrency’s risk,what are the crossmarket spillover effects of cryptocurrency’s volatility? How major emergencies affect these spillover effects? In terms of cryptocurrency’s portfolio strategies,how to leverage machine learning methods(spillover effects)to improve(reduce)the performance(risk)of portfolio strategies in the cryptocurrency market.This research conducted the following three comprehensive and detailed studies.First,this research uses machine learning methods to study cryptocurrency’s return predictions.This research constructs 8 macroeconomic factors and 18 characteristic factors from the macro and micro perspectives,respectively.This research comprehensively utilizes 12 machine learning models to predict cryptocurrencies’ returns in the full sample and market capitalization groups and evaluates the out-ofsample performance and differences of all models.This research employs the advanced SHAP(Shapley additive explanations)method to examine the importance of factors in the model and the impact of factors on model output.In addition,this research conducts a separate analysis focusing on Bitcoin’s return predictions based on its unique characteristics.The results show that tree models(especially random forest model)are the best performing machine learning method;tree models can better predict returns in the cryptocurrency market compared to the stock market.In the tree models,the ratio of market value to realized value is the most important factors.The greater the value of this factor,the greater the predicted return.Second,this research uses the time-varying parameter vector autoregression(TVP-VAR)model to study the price volatility spillover effects of cryptocurrencies and traditional financial assets.This research first utilizes the vector autoregression(VAR)model to examine the spillover effects in the cryptocurrency market,and then use the TVP-VAR model to explore the spillover effects of the cryptocurrency market and the traditional financial markets of the U.S.and China,respectively.In the studies based on the time-varying parameter model,this research not only examines the spillover effects between the cryptocurrency market and traditional financial markets,but also analyzes the spillover effects within the entire system and each market.In addition,this research also assesses the impact of spillover effects from two major emergencies on cryptocurrency markets and traditional financial asset markets.Results show that the cryptocurrency market has greater spillovers to traditional financial markets;compared with the total system spillover index of cryptocurrency and U.S.traditional financial assets,the fluctuation range of the total system spillover index of cryptocurrencies and China’s traditional financial assets is smaller.In addition,after the occurrence of emergencies,the spillover effects are significantly enhanced and their impacts are in short-term,and the connection between cryptocurrency and traditional financial markets becomes more closely linked.Finally,this research explores cryptocurrency investment strategies based on the research mentioned above.This research combines the 12 machine learning methods with 2 arbitrage strategies and 2 weight calculation methods to construct 48 machine learning portfolio strategies in the cryptocurrency market,and then evaluates the outof-sample performance of these strategies and their differences.Further,this research introduces a mean-variance model based on the performance-based regularization(PBR)and four mean-lower partial moment models based on PBR into the cryptocurrency market and examines the performance and differences of investment strategies under these mean-risk models in the cryptocurrency market.Besides,this research also employs tree model and volatility spillover effects to improve the mean-lower partial moment portfolio strategies based on PBR and analyzes the improvement of these new strategies.The results show that the machine learning portfolio strategy based on the neural network model performs the best;the mean-lower partial moment portfolio strategy that considers the PBR constraints of both risk and return measures performs the best;the performance of the mean-lower partial moment portfolio strategies based on RF model and considering volatility spillovers have been significantly improved.The innovation points of this article include the following three.First,this research comprehensively and systematically explores the pricing characteristics and price influencing factors of cryptocurrency.Second,this research fills the gap in research on volatility spillovers in the cryptocurrency and traditional financial markets.Third,this research introduces portfolio strategies based on machine learning methods in the cryptocurrency market and applies risk spillover effects to improve these strategies.