This study is devoted to compiling an investment portfolio based on general information about the client and a special questionnaire that determines the degree of risk appetite. The research includes a comparison of the effectiveness of various machine learning algorithms including linear regression, recurrent neural networks, decision trees, and Google Simple ML service. While most investment portfolio construction approaches are based on information about past assets price movements, the approach described in this paper proposes the application of machine learning algorithms to determine the shares of financial instruments’ types in the portfolio (e.g., Commodities, High Yield Bonds, etc.). The novelty of our approach lies in forecasting the shares of financial asset classes based on a real dataset of various characteristics of new investors, as opposed to existing studies that focus on rebalancing an already existing portfolio. The results of this investigation have practical applications in the automated generation of personalized investment portfolios in Robo-Advisor services.