IBM and Confluent bring foundation models for time series closer to streaming data. Natively integrated into Confluent, in the cloud or on-premises, these models, whose family includes four models and over 50 million downloads, run on CPU processors, without GPUs (graphics processing units).
Every company produces a continuous stream of observations: a machine measurement, a payment, a sale, an order, or a stock variation. These time series contain signals that can help anticipate what comes next. The challenge is to turn them into decisions while they still matter. A forecast delivered after a production run, a stock-out, or a payment authorization may be accurate, but arrive too late to influence the outcome.
Foundation models, already familiar for text, images, and documents, apply a similar idea to numerical data. They learn general trends from varied historical data, then apply this knowledge to new series. A starting point can thus be provided without training a separate model for each series. Depending on the task and deployment, teams can also fine-tune the model or combine its results with other data and business logic.
Concretely, the models run directly within Confluent as built-in functions. In Confluent Cloud, they are fully managed and require no deployment. On-premises, they can run in the company’s infrastructure via Confluent Platform for Apache Flink. Flink, a data stream processing engine, enables applying forecasting, anomaly detection, and semantic intelligence to continuous data, using SQL.
A Streamhouse for Real-Time Context
Companies operate continuously, while information is often spread across systems updated at different times, or depends on periodic batch processing. The Streamhouse aims to continuously capture, process, govern, and serve this current business context. Confluent’s Data Streaming Platform brings together five capabilities: connect, stream, process, govern, and query. Apache Flink transforms and enriches data as it flows, while Lightning Tables offer low-latency (very short response time) access to the current state.
Time series models are added to this architecture, alongside large language models, statistical models, and specialized machine learning systems. The use cases described in the article illustrate this combination. In an ice cream production line, quality depends on overrun, meaning the amount of air incorporated into the mix. A pre-trained TTM model reads SCADA data and quality control measurements every 10 minutes, and forecasts the share of compliant tubs four hours in advance, with a probable range. When this forecast falls below the factory’s target, an optimization process evaluates combinations of draw temperature and air flow rate. The operator approves the recommended setpoints.
In distribution, the AI_FORECAST function produces, in the example presented, a 28-day demand forecast with a P10–P90 range, updated as new sales arrive. A planning application calculates a reorder point, and an agent can prepare a purchase order for validation. In banking, each payment is compared to an expected spending range based on the card’s history. A transaction outside this range is forwarded to existing fraud systems, which can approve the payment, request additional authentication, or initiate further review. For trading, the PatchTST-FM model forecasts the trading volume for the next 30 minutes at each one-minute candle.
In every sector, the Streamhouse does the same work: Flink executes Granite models to forecast, detect, and optimize, streaming agents act on the live context, and people validate exceptions.
