Generating data for predicting court decisions in Kazakhstan using machine learning

Abstract

This study presents the development of a synthetic dataset and machine learning models for predicting court decisions in Kazakhstan. The dataset contains 100,000 cases generated from the Code of the Republic of Kazakhstan, covering both administrative and criminal offenses. Each record includes attributes such as the age of the accused, offense type and severity, and mitigating or aggravating factors. Regression models were applied to estimate offense severity, level of guilt, and likelihood of penalties, while classification models predicted the offense category, relevant law articles, and sentencing type. Predictions addressed both general outcomes—classifying cases as criminal or administrative—and specific judicial decisions, including fines, imprisonment terms, and other penalties. Classification models achieved 92% accuracy in determining offense category and sentencing type, and regression models reached a root mean squared error (RMSE) of 0.12 for offense severity. Using synthetic data preserves confidentiality while enabling pattern discovery for decision support. The results demonstrate the potential of artificial intelligence (AI) to improve sentencing prediction, prioritize case processing, and enhance transparency in Kazakhstan’s judicial system. Beyond transparency in decision support, the proposed approach also shows potential in crime prevention, workload optimization, and fostering digital transformation within judicial operations.