| Asset allocation is one of the most popular topics for asset management in modern society.However,from personal and household finance to governments and enterprises’ investment decisions,all the investors would need some scientific methods to assist their investment.Therefore,this paper would focus on the topic of major asset allocation,which involves three submodules,i.e.,major asset timing,assets’ returns distribution prediction and their application in portfolio optimization.The detailed works can be described as follows:(1)Machine Learning-based Major Asset Investment TimingThe purpose of investment timing is to predict the trend of assets’ prices,i.e.,whether their prices will rise or fall.This is actually a binary classification problem.By summarizing related literatures,this paper would firstly select a group of technical indicators which may have significant influence on the prediction task.Secondly,with the training subsets generated by Bagging algorithm,numerous different and diversified extreme learning machine models could be trained as basic classifiers.Finally,the prediction of each basic classifier would be integrated into the final prediction results,via the deep learning technique.The experimental results show that this novel model beats both typical single classifiers and ensemble models in several tested scenarios.(2)Decomposition and Ensemble-based Assets’ Returns Distribution PredictionThe“time series decomposition and ensemble”principle is firstly introduced to the task of returns distribution prediction,in which three sub-steps are involved:data decomposition,individual returns distribution prediction,and prediction results integration.First,the ensemble empirical mode decomposition technique is employed to decompose financial assets’ price time series into several independent time series.Then,the quantile regression neural network model is used to predict the distribution of each decomposed time series.Thirdly,the prediction results of each decomposed time series are integrated via a simple addition method.The simulation study on the Composite Index of Shanghai Stock Exchange shows that the proposed ensemble learning paradigm can significantly outperform its counterpart benchmarks,and also the historical returns’ statistical method.(3)Investment Timing and Returns Distribution Prediction-based Portfolio OptimizationIn terms of the final portfolio optimization,the timing model is firstly used to filter out the assets which are predicted to fall.Then,based on assets’ joint returns distribution modelled by copula function and their marginal returns distribution prediction results,a simulated returns data can be generated via the Monte Carlo method.Finally,classic portfolio models are optimized on the simulated data,and thus the optimal investment decisions for major asset class are obtained.Through experimenting with real-world market data,we could conclude that the returns distribution prediction-based portfolio model is effective and powerful.The three modules proposed in this paper are interrelated and complemented with each other,which constitute a complete asset allocation methodology.In the empirical study,this methodology is verified by the real market data.The results show that the proposed major asset allocation model achieves better investment performance comparing with its benchmark models,in terms of Sharpe ratio and accumulated return. |
