Real-Time Intelligence with IBM Time Series Models on Confluent

Foundation models have changed how enterprises extract value from unstructured data, but the larger opportunity lies in streaming data, where mission-critical decisions—such as demand forecasting, payment holds, and equipment maintenance—are made. Until now, these decisions relied on bespoke models built one at a time over months of expert work, forcing teams to model only a few hundred high-value series and cover the rest with costly safety margins that are applied after the decision window has closed.

IBM and Confluent are addressing this by bringing time series foundation models (TSFMs) directly to the data stream. Trained on vast and varied signals, a TSFM generalizes to previously unseen series, providing forecasts, anomaly scores, historical analog matches, and suggested settings for target outcomes. The models are designed to be used by domain experts—demand planners, fraud analysts, process engineers—without requiring a dedicated data science team. IBM is packaging these capabilities, including forecasting, anomaly detection, optimization, and semantic intelligence, as callable functions rather than custom-built projects.

As an example, on a chocolate factory tempering line where temperature, speed, and throughput are sampled every few seconds and monitored against fixed thresholds, a foundation model can forecast line output through the evening shift, flag slow drifts relative to normal dark-chocolate behavior, and find the closest matching historical production run. It conditions on controllable settings and can fine-tune when higher accuracy is needed, all with the same model applicable across every line and factory.

IBM validated these models first internally and then with design partners in cement, steel, pulp and paper, food, and telecommunications, reporting that every point of accuracy is worth millions, productivity gains run 5–10×, and work once delegated to specialists now sits with the domain experts who own decisions. The models are now live in Early Access on Confluent Cloud, hosted where the data moves and callable from Flink, with 44M+ downloads behind IBM's time series models. Confluent Platform support is planned next.