An excessively high Gas-Oil Ratio (GOR) can reduce the cumulative oil production of individual oil wells during CO2 flooding. Therefore, predicting the maximum GOR to extract prevention measures holds paramount significance. Present methods for predicting the maximum GOR, predominantly employing numerical simulations and empirical formulas. However, numerical simulations often overidealize the real problem and empirical formulas result in inaccurate predictions. This paper introduces an Improved Random Forest (IRF) algorithms for maximum GOR forecasting accurately. This study introduces an Improved Random Forest (IRF) algorithm, to enhance the precision and generalization capabilities in predicting the maximum Gas-Oil Ratio (GOR). The IRF algorithm integrates three significant modifications into the traditional Random Forest (RF) framework: pre-classification of samples, stratified random sampling, and optimized voting weights for decision trees. These modifications theoretically offer dual benefits: enhanced extraction of information from underrepresented classes, and a more reasonable weight of decision trees in the model, thus promising a more robust and accurate prediction of maximum GOR. 264 data samples were collected from typical CO2 flooding models with nine characteristics. Results from 30 repeated experiments under consistent conditions demonstrated that the IRF algorithm outperformed the conventional RF algorithm. The RF algorithm yielded an average coefficient of determination (R2) of 0.71, alongside a higher Mean Absolute Error (MAE) of 603.2. The IRF algorithm achieved an average R2 of 0.83. The MAE was notably reduced to 361.9. The enhanced IRF algorithm demonstrated a 12.1% improvement in accuracy, an 2.2% increase in generalization ability, and a significant 40.00% reduction in MAE. The substantial decrease in MAE indicates the algorithm’s effectiveness in reducing model generalization error, reflecting enhanced performance on unseen data. Thus, it's evident that in predicting the maximum GOR, the IRF surpasses the standard RF in both predictive accuracy and generalization capability. Stratified random sampling, combined with the adjustment of the decision trees’ voting weights, is instrumental in significantly enhancing the precision of the predictions. The innovation of the IRF algorithm enables precise prediction of the maximum GOR for individual oil wells, tailored to the unique characteristics of different oil reservoirs. This model provides a highly effective tool for predicting maximum GOR, thereby facilitating the management of gas channelling in varied reservoir condition and diminishing the reliance on reservoir engineers’ expertise.