The purpose of this study is to quantitatively research investment risks in various countries worldwide, generate risk indicators and their standard deviations, and provide references for enterprise investment decisions. In contrast to prevailing methods that involve manually selecting risk indicators and assigning subjective values, this study utilizes big data and machine learning to control significant flaws such as selection bias and subjective weighting in the process of generating risk indicators.

The research team collected and organized data from a vast array of sources, amassing over 8,000 pieces of data. Each data point includes information spanning over 200 dimensions, covering various aspects of geography, history, culture, economy, politics, and society for 151 countries over 54 years. The scale, dimensionality, comprehensiveness, and coverage of the information significantly exceed existing risk assessment systems.

In terms of statistical methods, the study employed multiple imputation to address data missingness. Various statistical techniques, including correlation analysis and univariate regression models, were used for data dimensionality reduction. Machine learning methods, specifically the LASSO model, were applied to select risk indicators. After determining the variables through machine learning, the study established predictive models for political risk and policy risk based on the preceding one, two, and five years.