Salesforce builds self-evolving forecast drift agent

salesforce.com5 min read

Salesforce has developed an internal self-evolving data science agent to autonomously investigate and fix forecast model drift across its infrastructure platform. The agent tests candidate model fixes, such as swapping forecasting algorithms or tuning parameters, against historical data using Kubernetes and Spark, and then opens a pull request for human data scientists to review. To date, the agent has proposed over 100 model fixes, reducing sMAPE error by 14% in backtests and cutting investigation times from hours to 27 minutes. Open article →