In supply chain management, the accuracy of demand forecasting is essential to optimize inventory and improve operational efficiency. However, there are often outliers in the demand data (such as sudden demand changes or human error) that, if not effectively addressed, can seriously interfere with the forecasting model. This study discusses how to improve the accuracy of supply chain demand forecasts by anomaly detection and treatment. The IQR is used to detect and smooth the abnormal demand data, and the demand prediction model is used to correct it. The experimental results show that the accuracy of demand prediction after smoothing is improved, especially when the demand changes and fluctuations are captured. KPI analysis shows that the prediction model after anomaly detection is better than the unprocessed data on RMSE and MAE indicators. This method provides an effective solution for demand forecasting for dealing with large fluctuations or abnormal data and can optimize supply chain management decisions.

Since the advent of artificial intelligence, it has not only transformed the way we live, but is also accelerating the transformation of production methods. Machine learning is a pivotal technology that endows computers with intelligence. The primary driving force behind its advancement is the pursuit of rapid and precise knowledge acquisition. Nevertheless, the existing enterprise risk decision support system is inadequate in terms of both timeliness and effectiveness when confronted with the task of analysing vast quantities of data. Furthermore, it lacks the capacity to assimilate the expertise of managers and to facilitate interaction in an intuitive manner.By combining a variety of machine learning models, the ensemble learning method effectively leverages the advantages of different algorithms, thereby improving the accuracy and stability of risk prediction. Each model provides a unique perspective and decision-making boundaries when dealing with complex enterprise risk data, but a single model may not be adequate when faced with a particular data set or a particular type of risk. The ensemble method overcomes the problem of bias or overfitting that may be caused by a single model by combining the prediction results of multiple models, so as to obtain a more robust prediction effect. These methods optimize the final prediction results based on how well they perform during training by giving different weights to each model. This convergence strategy significantly improves the accuracy of enterprise risk assessment and helps to provide decision-makers with more reliable data support to make more informed risk management decisions in complex business environments.