Real-Time Intelligence with IBM Time Series Models on Confluent

Foundation models have changed how enterprises work with unstructured data. The next opportunity is streaming data, where many operational decisions occur: how much to order, which payment to stop, when equipment may fail, how to run a production line, and whether a current event resembles a previous one.

IBM and Confluent are making time series foundation models available in Early Access on Confluent Cloud. The models run where streaming data already moves, with Confluent Platform planned to follow.

Why time series foundation models matter

Many organizations have traditionally built one bespoke model at a time, requiring months of specialist work. As a result, teams often model only the few hundred time series with the clearest financial value. Other series are managed with safety margins, such as additional inventory, capacity, or operating tolerance. Those margins represent the cost of decisions made without reliable forecasts.

A time series foundation model (TSFM) is trained across large and varied collections of signals. It can generalize to a series it has not previously encountered. Given a window of measurements, it can forecast future values, identify deviations from normal behavior, find similar historical patterns, and evaluate settings against a target.

These capabilities can be used by domain specialists, including demand planners, fraud analysts, and process engineers. IBM is also developing functions around the models so forecasting, anomaly detection, optimization, and semantic intelligence can be accessed as platform capabilities rather than built as separate projects.

Consider a chocolate tempering line whose temperature, speed, and throughput are sampled every few seconds. A foundation model could forecast output through the evening shift, allowing a planner to identify a likely shortfall while there is still time to respond. It could compare the current run with the line's normal behavior for dark chocolate, detect gradual drift before it affects the product, and identify similar historical runs and their outcomes. The model can also account for operator-controlled settings and be fine-tuned when additional accuracy is worthwhile.

IBM tested these models in its own products and operations before working with design partners in cement, steel, pulp and paper, food, and telecommunications. The reported results include accuracy improvements valued at millions of dollars, productivity gains of 5 to 10×, and the transfer of work from specialist teams to the domain experts responsible for the decisions.

IBM contributes time series models with more than 44M downloads, while Confluent provides the live business context and connections to downstream systems. The models run in a stream-native architecture, hosted in Confluent Cloud and called from Apache Flink. Access begins on Confluent Cloud on AWS, with Confluent Platform planned for on-premises and hybrid environments.

Time series intelligence and real-time context

The time required to put a model into production can be significant. In this architecture, IBM Granite reads the signal while Confluent supplies the streaming context, governance, and delivery to downstream systems.

The value of a signal can decline quickly. A pump identified as drifting today may generate a maintenance work order, while the same discovery next week could correspond to an outage.

Forecasting and anomaly detection are stateful operations. A future value is meaningful in relation to recent history, and an anomaly depends on a continuously updated understanding of normal behavior. Flink manages state by series and provides fault tolerance, allowing each model to access the history it needs without a separate data store or a database request for every inference call.

Confluent's data streaming platform continuously connects, governs, and processes real-time data. It captures signals from sources such as sensor telemetry, payment activity, and application metrics, making them available to IBM Granite Time Series models for:

Forecasting
Anomaly detection
Similarity search
Classification
Gap-filling
Optimization

Confluent Cloud provides native inference, allowing IBM Granite Time Series models to run directly within Apache Flink on Confluent. This unifies data and machine learning workflows for real-time processing.

Key capabilities include:

Real-time intelligence where data lives: Forecasting and anomaly detection run directly on streaming data, without moving time series into a separate machine learning platform or data warehouse.
Zero configuration: Confluent manages model serving, infrastructure, scaling, and runtime operations. Models can be called from Flink SQL using the AI_FORECAST and AI_DETECT_ANOMALIES functions.
Fresh, enriched context: Models use an up-to-date view of business conditions rather than stale batch data. Inference results are written to Kafka topics and can be consumed by alerting systems, dashboards, lakehouses, and AI agents.
Governance and traceability: Inference pipelines use the platform's schemas, lineage, and access controls. Durable, replayable Kafka topics support auditing, troubleshooting, model evaluation, and rerunning inference on historical data.
Cost efficiency: Native inference avoids the need to provision and manage dedicated model-serving infrastructure or GPUs, and does not incur cloud ingress or egress fees.
Security: Data remains within Confluent Cloud during inference and continues to follow role-based access control and privacy policies.
Faster implementation: Teams can create forecasting and anomaly-detection pipelines using familiar SQL syntax without building a separate machine learning stack or point-to-point data pipelines.

Confluent connects operational and analytical environments so live events can become actionable intelligence. Because different applications require different model characteristics, IBM and Confluent provide a portfolio rather than a single model.

A portfolio of time series foundation models

Different decisions require different forms of time series intelligence. A planning process may need a range of outcomes, while a trading application may prioritize accuracy across multiple sampling rates. A fleet with hundreds of thousands of series may require low inference costs, while a security team may need to detect behavioral changes and identify similar historical events.

The portfolio includes four complementary time series foundation models. They are available in Early Access through Confluent's existing AI_FORECAST and AI_DETECT_ANOMALIES Flink SQL functions. The model can be changed with one SQL parameter, without redesigning the pipeline.

SELECT
  AI_FORECAST(
    load_kw,
    event_time,
    JSON_OBJECT('model' VALUE 'ttm', 'horizon' VALUE 12)
  ) OVER (
    ORDER BY event_time
    RANGE BETWEEN UNBOUNDED PRECEDING AND CURRENT ROW
  ) AS forecast
FROM meter_readings;

Changing the model value allows the same call to use any of the four models.

PatchTST-FM: Processes a series in patches, with each variable handled in its own channel. It returns a full distribution, allowing a planner to set policies such as reorder points based on the 90th percentile.
FlowState: Maintains a running summary updated with each data point. Its continuous-time dynamics support data ranging from seconds-level SCADA measurements to hourly market observations.
TTM: Uses small mixing networks along time and across variables rather than attention. Its approximately million-parameter design can cover 100,000 series nightly on CPU.
TSPulse: Combines time-domain and frequency-domain views in a compact multi-task model for anomaly detection, classification, gap-filling, and similarity queries.

The models are designed to run natively inside Confluent Cloud or on CPUs using open weights from the Hugging Face Hub. The architecture does not require cloud ingress or egress. IBM Granite also includes IBM's enterprise AI governance framework, model provenance, licensing transparency, and functions designed for specific time series use cases.

Applications across the enterprise

The four models reduce the time between an event and the point at which an organization understands it. In a streaming environment, that interval can shrink from days to seconds. The resulting signals can trigger AI systems, agents, and workflows that investigate events, prioritize them, and involve a person when necessary.

Forecasting and planning

Traditional forecasting often combines statistical models with manual judgment. Bespoke machine learning can require one model per series, manual refitting, and continued maintenance as forecast horizons change. This limits detailed planning to the series that justify the effort, while other items are covered by safety stock.

A demand planner at a grocery retailer could apply one shared model across an entire catalogue. The model can work on an unseen series, incorporate drivers such as weather and promotions, and return a probability distribution instead of a single forecast line.

This approach supports catalogues containing both long-established and recently introduced products. A new item with little history can be associated with similar products, while up to 100,000 SKUs can be forecast nightly on CPU across categories and regions.

A forecast distribution can translate service-level requirements into explicit policies. Because each forecast is published to a topic, it can also trigger replenishment, allocation, pricing, and markdown workflows. The intended business effects include fewer stockouts and markdowns, improved product availability, protected revenue, and reduced working capital.

Anomaly detection

Anomaly detection applies to fraud, security, IT operations, and industrial systems. A missed anomaly can involve a customer's money, a breach, or an outage. Rule-based systems can be enumerated by adversaries, while stricter rules can reject legitimate activity. Bespoke machine learning models may also depend on labels that are limited or become outdated, and excessive false alarms carry their own costs.

A retail bank's fraud team could maintain a model of normal behavior for each card and score every payment through the AI_DETECT_ANOMALIES function while the transaction is in progress. Because the system also forecasts behavior, it can identify movement toward an abnormal event before the event occurs.

For example, a card that has purchased groceries in the same three postal codes for two years might suddenly fund a wallet abroad at 3 a.m. The system could raise an alert before funds move, while allowing a legitimate customer on holiday to proceed when the broader behavior remains consistent.

The models can transfer learned behavior to new products, corridors, and asset types without requiring an initially labeled case. They can also be customized using a bank's own stream and updated as cases are confirmed. This is important because fraud patterns and attack signatures change over time.

The same approach can apply to any entity with a recurring pattern, including IT latency, cell-site KPIs, and manufacturing equipment.

Production optimization

Industrial plants rely on models of their own processes, but those models can drift. Rule-based controls maintain setpoints without necessarily improving them, and bespoke machine learning may be difficult to explain. These limitations can keep optimization projects confined to pilots.

A process engineer at a shampoo plant could stream temperature, agitator speed, dosing rate, and viscosity into Confluent. A foundation model can begin working without months of bespoke modeling, while customization is reserved for requirements specific to the production line. When conditioned on controllable settings, the forecast becomes a simulator for questions such as:

What energy use results from a particular mixing speed?
What throughput is expected at a given temperature and dosing rate?
Will viscosity remain within specification?

An optimizer can search the operating space against a named KPI while respecting constraints. Recommendations can be linked to the underlying conditions and historical examples, helping process engineers evaluate them.

Optimization can continue as inputs change. Supplier changes, different fragrance batches, shifts in demand, and new production objectives may all require updated settings. Restating the objective and constraints allows the system to identify another suitable operating point, starting from similar historical runs and their corrective actions.

The financial impact can be significant. A one-point improvement in an operation generating hundreds of millions of dollars can represent a seven-figure amount. One food manufacturer can begin with one process and extend the approach across 400 factories.

Production optimization can also support quality prediction and equipment-condition monitoring as AI expands into manufacturing, robotics, and other physical systems.

Semantic intelligence

The applications above often lead to the same question: has this happened before? Time series embeddings address this by representing the shape of a time and frequency window as a compact vector. Similar episodes can therefore be matched even when their scale or offset differs.

As each window arrives, it can be embedded and compared with historical episodes and their outcomes. The result is a precedent rather than only a score: previous production runs that drifted in the same way, demand curves resembling a new product, or confirmed fraud cases similar to a current session.

The same embeddings can support classification and gap-filling, as well as indexing context for an agent to retrieve before taking action.

Early Access on Confluent

IBM Granite Time Series models are available in Early Access on Confluent Cloud for forecasting and anomaly detection on streaming data. The offering is designed to operate without separate model training, feature engineering, or specialized AI and machine learning expertise. Confluent Platform is planned as the next deployment option for on-premises and hybrid environments.

The Early Access functions include:

AI_FORECAST for real-time forecasting
AI_DETECT_ANOMALIES for real-time anomaly detection

Documentation for both functions is available through Confluent Cloud for Apache Flink. The models and their capabilities are being developed around enterprise use cases involving productivity, accuracy, and responsiveness across forecasting, anomaly detection, optimization, and semantic intelligence.

Real-Time Intelligence with IBM Time Series Models on Confluent