Research On Mutual Fund Portfolio Allocation Of Robo-advisor Based On Deep Learning

With the continuous growth of residents’ wealth management needs,the investment advisory business has entered a golden stage of development,since human advisors fail to provide more abundant and efficient financial services,robo-advisors emerged as a result.Robo-advisors can enhance the digitalization of financial institutions,optimize resource allocation and expand service scope,provide technological innovation for the traditional financial services,and drive the development of the service models in the realworld economy,which has received widespread attention from the financial industry and academia.Most of the existing research focuses on the business model,policy,and supervision of robo-advisors,while few literatures have conducted in-depth research on its investment decision-making process,not to mention the relevant research on the application of artificial intelligence.This paper takes actively managed mutual funds as the research object,based on modern portfolio theory,behavioral finance theory,and economic cycle theory,uses deep learning methods to construct a research framework for robo-advisor’s fund portfolio allocation,and discusses the application and result of machine learning in fund classification,deep learning in fund evaluation,and economic cycle analysis in fund portfolio construction,hence,provide a theoretical and practical basis for promoting the better integration of artificial intelligence and mutual fund portfolio allocation,and promoting the better development of robo-advisory business.This paper takes fund classification-fund evaluation-fund portfolio construction as the mainline of research.First,mutual fund classification using feature data.Through the analysis of fund features and the study of fund classification methods,the classification goal is determined and the feature data is preprocessed.A Gaussian mixture clustering model was used to classify funds in two steps.Second,building a cluster-based deep learning ensemble model to predict fund performance.The virtual net asset value of fund is constructed using the holding information released in quarterly reports to track fund movement and detect significant trading activity by fund managers.On this basis,a deep integrated learning model based on Residual Network(Res Net),Long Short-Term Memory Neural Network(LSTM),and Convolutional Neural Network(CNN)is constructed using the historical net asset value to enhance the predicting ability on the future performance of the fund.Third,using multi-dimensional fund evaluation indicators to predict the future performance of funds under each category using a spatio-temporal deep learning model.The evaluation indicators of the fund manager’s management abilities have been added to the fund evaluation system.On this basis,a deep learning model based on Conv LSTM,Res Net,and fully connected neural network layers(FC)is constructed,and the future excess return capability of the fund through multi-dimensional fund evaluation indicators is predicted.Fourth,Construct a macro-economic scenario for robo-advisors’ active asset allocation,and construct optimal fund portfolios based on economic state rotation.Currency-economy-price boom index is used to divide China’s historical economic cycle into six states.On this basis,a Transformer-Matrix deep learning model is constructed to predict the transitional probability between different economic states.Finally,through scenario analysis,the fund categories with the best historical performance in each economic state are identified,the funds with the highest evaluation scores under each category are selected,and the dynamic allocation of the optimal fund portfolio is carried out using the economic state transitional matrix.Main conclusions of the paper:(1)The fund classification method based on two-step Gaussian mixture clustering can achieve similar accuracy to the traditional post-positionbased classification method without knowing position information,with higher timeliness.The model classifies market-wide mutual funds into ten parent classes through the first step,and divides the fixed income and hybrid funds into three subclasses through the second step.(2)Tracking the deviation of the actual net asset value through the virtual net asset value can capture the fund manager’s trading activity in time,and the virtual net asset value has a better tracking and prediction effect on the fund manager who holds a big position in a long-term manner.Predicting the future performance of the fund through the deep ensembled learning model tracks benchmark better than the virtual net value using the holding information,and can be used as an alternative to the real-time monitoring and short-term evaluation of the fund,especially when the fund manager who makes a big adjustment to the fund positions during reporting periods.(3)With the addition of evaluation indicators on fund managers’ management abilities,the dimensions of the fund evaluation system have been expanded,reducing the correlation between indicators.Judging from the results of evaluation indicators,actively managed funds consistently generate excess returns over the benchmark,mainly because of fund managers’ positive sector allocation capabilities and active management skills.The performance of top-rated funds in each category based on spatio-temporal deep learning model is significantly better than that of traditional multi-factor models,mainly because deep learning models can effectively extract nonlinear characteristics between different fund evaluation indicators.(4)Using the data of the past 15 years,under monthly observation,the probability of transition between different economic states follows an orderly fashion with the highest probability of self-rotating.Under the economic cycle scenario,the performance of different types of funds is analyzed,and it is found that the small and medium-cap balanced style in the early stage of prosperity is dominant,the consumer discretionary and healthcare thematic fund in the late boom period is dominant,the large-cap value style is dominant during the stagnation period,and the consumer discretionary and healthcare thematic fund re-dominant during the recession period.The optimal fund portfolio based on the economic state rotation has significant excess returns under monthly and quarterly rebalance compared with the portfolios holding all funds or the mainstream fund index in the market.Performance attribution shows that the optimal portfolio has more exposure in momentum and growth factor than the CSI 300 Index,while having less exposure in the size factor.Major innovations of the paper:(1)Established theoretical and empirical research and analysis framework for robo-advisor fund portfolio allocation.Theoretically,further deepens the integration of modern portfolio theory with behavioral finance and economic cycle theory,and expands the scope of theoretical application.Empirically,the optimal fund portfolio construction under machine learning and deep learning methods is studied progressively from the fund classification,evaluation,and portfolio allocation.(2)Applied and verified machine learning and deep learning in the decision-making of specific financial problems.The spatio-temporal characteristics and nonlinear correlation relationships between fund data are fully obtained to predict the future fund performance by introducing Res Net,LSTM,Conv LSTM,FC,Transformer,and other deep learning models.At the same time,in terms of deep learning methods,the idea of clustering and ensemble learning is used to effectively improve the generalization and prediction abilities of the model.(3)Introduced economic cycle analysis into fund portfolio construction of robo-advisors,achieving the goal of active asset allocation through scenario analysis,which provides a new idea for the portfolio construction of roboadvisors.Through the comparison with the mainstream fund portfolio index in the market,portfolio allocation of robo-advisory funds under scenario analysis of economic cycles can obtain excess returns,making up for the lack of domestic practice on robo-advisor asset allocation.