BMW Group has added daily anomaly detection to the system that monitors cloud spending across more than 14,000 accounts. The internal Cloud Efficiency Analytics platform now forecasts expected costs and emails account owners when a service’s spending moves far enough outside its normal pattern.
The platform ingests about three billion billing rows with 500 columns each month, then aggregates them into daily costs for every account and service. Each of those hundreds of thousands of time series receives a forecast based on 365 days of history. AWS Step Functions distributes the work across as many as 500 Lambda functions, completing the full run in roughly 20 minutes for about $50 a month in compute.
A statistical deviation alone does not trigger an email. The system excludes very small services and those with less than ten days of history, requires a move of at least 40 percent and applies minimum dollar impacts based on the account’s spending group. Teams can add exceptions for services that are volatile by design.
BMW uses Meta’s open-source Prophet forecasting library, but designed the model component to be replaceable. One limitation is that an adaptive forecast can absorb a sustained increase after flagging its first days, making the method strongest at finding spikes rather than every lasting cost change. Human owners still decide whether an alert represents waste or planned growth.