IBM Granite Time Series Models on Confluent

IBM and Confluent have integrated IBM Granite Time Series foundation models into Confluent Cloud, allowing enterprises to perform real-time forecasting and anomaly detection directly on streaming data. This integration eliminates the need for bespoke model development for every data series and removes the requirement to move data into separate ML platforms or data warehouses.

Stream-Native Intelligence with Apache Flink

IBM Granite Time Series models are now available in Early Access on Confluent Cloud, running natively within Apache Flink. This architecture allows models to operate where the data moves, providing several operational advantages:

  • Zero Configuration: Model serving, scaling, and runtime operations are managed by Confluent, allowing users to call models directly from Flink SQL without managing provider credentials or glue code.
  • Stateful Processing: Because forecasting and anomaly detection require recent history to define "normal" behavior, Flink manages the necessary state per series in a fault-tolerant manner, removing the need for separate database hits per call.
  • Built-in Governance: Inference pipelines follow the same schemas, lineage, and access controls as the rest of the Confluent platform. Kafka topics ensure that results are durable and replayable for auditing and model evaluation.
  • Cost and Security Efficiency: Native inference eliminates the need for dedicated GPU infrastructure and removes cloud ingress/egress fees, while keeping data within Confluent Cloud's RBAC and privacy policies.

The Time Series Foundation Model Portfolio

Rather than a single model, IBM provides a portfolio of four complementary foundation models designed for different decision-making needs. These are accessible via the AI_FORECAST and AI_DETECT_ANOMALIES Flink SQL functions.

Model Specializations

  • PatchTST-FM: Processes series in patches (similar to how LLMs read text) with each variable in its own channel to prevent noisy signals from impacting others. It provides a full distribution of outcomes, enabling planners to set thresholds based on percentiles (e.g., the 90th percentile).
  • FlowState: Maintains a running summary updated with every data point. Its continuous-time dynamics allow it to handle varying frequencies, from seconds-level SCADA data to hourly market data.
  • TTM: Utilizes tiny mixing networks across time and variables instead of attention mechanisms. This allows a million-parameter model to cover hundreds of thousands of series nightly on CPU.
  • TSPulse: A small multi-task model combining time and frequency views, specifically optimized for anomaly detection, classification, gap-filling, and similarity search ("have we seen this before").

Enterprise Application Lanes

The integration focuses on four primary technical capabilities that compress the time between a business event and an actionable response.

Forecasting and Planning

Foundation models allow for a "model factory" approach where a single shared model works on unseen series immediately. This is particularly effective for the "tail" of a product catalogue where bespoke models are too expensive to build. By outputting a distribution rather than a single line, these models enable more precise service-level policies, reducing stockouts and freed working capital.

Anomaly Detection

By maintaining a sense of "normal" per entity (such as a credit card or a sensor), the models can score payments or telemetry in-flight. Because the models are foundation models, they provide day-one protection for new products or asset types without requiring labeled historical cases. They can be continuously refit on a company's own stream to adapt to evolving fraud patterns or attack signatures.

Production Optimization

In industrial settings, these models act as simulators. By conditioning on controllable variables (e.g., temperature, speed), engineers can forecast outcomes and use an optimizer to search for the best setpoints to meet a specific KPI. IBM reports that this approach can lead to productivity gains of 5x to 10x compared to traditional bespoke modeling.

Semantic Intelligence

The models utilize embeddings—compact vectors capturing the shape of a window in time and frequency. These embeddings allow the system to match current streaming windows against past episodes to find precedents. This enables the retrieval of historical fixes for similar drifts or the identification of similar fraud cases, providing context for AI agents before they act.

Sources