Models, Standard

Time-Series Models in the Streamhouse: Real-Time Forecasting on Streaming Data

A glowing shared platform absorbs a river of streaming data and projects bright forecast arcs forward, while cracked isolated silos crumble on the far bank

The time-series models Streamhouse integration, announced by IBM Research on Hugging Face on 18 November 2026, brings forecasting, anomaly detection and semantic search to streaming data. IBM and Confluent have put IBM Granite time-series foundation models natively inside Confluent, where Apache Flink can call them with SQL — so predictions are made on data while it is still flowing, not after it has been warehoused. The models have been downloaded more than 50 million times.

The timing problem the launch addresses is simple. As the announcement puts it: “A forecast delivered after a production run, a stockout, or a payment authorization may be accurate, but too late to influence the outcome.” Every business generates a continuous record of change — a machine reading, a payment, a sale, an order, a change in inventory — and those observations carry signals about what happens next. The point is timing: a forecast only helps the decision at hand if it reaches the people and applications responsible for it while they can still act on it.

What a Streamhouse is — and why agents need current context

A Streamhouse is Confluent’s answer to a familiar problem: business events flow continuously, but the information behind them often sits across systems that update at different times or rely on periodic batch processes — so applications and AI agents can end up acting on information that no longer reflects what is happening.

A Streamhouse addresses this by continuously capturing, processing, governing and serving current business context. Confluent’s Data Streaming Platform brings together five capabilities — connect, stream, process, govern and query — to make that context available to applications, analytics and agents; Apache Flink transforms and enriches data as it flows, and Lightning Tables provide low-latency access to the current state.

Time-series models add predictive power to this architecture. Flink can prepare the data, invoke forecasting and anomaly detection, and make the results available alongside the events that produced them; other models — large language models, statistical models and specialised ML systems — can contribute where appropriate. The pattern is the same in every industry: “Flink runs the Granite models to forecast, detect and optimize, Streaming Agents act on the live context, and people approve the exceptions.”

What time-series foundation models are

Four distinct glowing model shapes — a crystal orb, a wave, a cube of light and a pulsing sphere — float above a timeline ribbon of data points

Foundation models have changed how organisations work with text, images and documents. Time-series foundation models extend a similar idea to numerical data: they learn general patterns from diverse historical data and apply that knowledge to new series and tasks, providing a starting point without requiring teams to train a separate model for every series. Depending on the task and deployment, teams can also fine-tune a model or combine its outputs with other data and business logic.

The IBM family behind the launch comprises four complementary models, each matched to a different decision:

  • PatchTST-FM — the all-around performer, strongest at probabilistic forecasting with full distributions and quantiles.
  • FlowState — sampling-rate invariant, with the best point-forecast accuracy.
  • TTM (Tiny Time Mixer) — the best efficiency and performance trade-off; it supports control variables and runs many series at low cost on CPU.
  • TSPulse — a 1M-parameter model for anomaly detection, classification, similarity search and gap-filling.

The models are called from Flink through the documented AI_FORECAST and AI_DETECT_ANOMALIES functions; which models and versions those functions expose remains to be confirmed.

Time-series models Streamhouse: how they run natively in Confluent

The models run natively inside Confluent as built-in functions, in the cloud and on premises. In Confluent Cloud they are fully managed, with nothing to deploy; on premises, they run within the customer’s own infrastructure using Confluent Platform for Apache Flink. The models run on CPUs, without requiring GPUs — small enough to run on standard CPUs alongside the streaming pipeline, which can reduce the need for dedicated GPU infrastructure.

Teams invoke the models from Apache Flink with SQL, through two functions: AI_FORECAST for predicting future values and AI_DETECT_ANOMALIES for identifying unexpected patterns. Behind them are the time-series foundation models, whose advantage is zero-shot inference — the models can be applied to a series they have not been trained on, without standing up a separate machine learning environment.

Confluent manages the serving infrastructure, so there is no model server to deploy or operate. The September announcement described the launch as Early Access on Confluent Cloud on AWS, with Confluent Platform support for on-premises and hybrid environments planned next; the November post describes on-premises use via Confluent Platform for Apache Flink. The difference is flagged in the limits section below.

Forecasting, anomaly detection and semantic search on live streams

The models add three capabilities to streaming data. Forecasting predicts what comes next. Anomaly detection compares each new reading with an expected range based on the series’ recent history, at a confidence level the team chooses — without training a separate model for every series. Semantic search uses TSPulse to represent time-series windows so an engineer can retrieve historical patterns shaped like the one under investigation: a spike and recovery, a gradual rise, a step change.

Three things change when these capabilities run inside the stream. First, model development scales: one pretrained model provides a common starting point across many time series. Second, predictions refresh as data changes, incorporating recent observations as part of a streaming workflow. Third, signals sit closer to decisions: forecasts and anomalies reach operational applications sooner, where teams combine them with business context, rules and human judgment.

Semantic search carries one explicit limit: similarity can help narrow an investigation, but it does not establish a root cause by itself. The announcement describes this as a separately built application, where the representations are indexed and the engineer consults the associated context.

Six use cases, from factory floors to trading desks

A luminous stream of data splits into six branches leading to a factory, a store, a bank, a trading spire, a sensor tower and a search lens

The announcement works through six scenarios that show how organisations could combine time-series models, streaming data and application logic.

Production optimisation. The announcement’s example uses a pretrained TTM model that reads SCADA telemetry and QA measurements every 10 minutes and forecasts the share of tubs expected to be in spec four hours ahead, along with a likely range. TTM also supports control variables, so process settings can be part of the model’s inputs. This is the production-optimisation use case: a forecast of product quality, available while there is still time to act on it.

Anomaly detection. In the announcement’s store-sales example, AI_DETECT_ANOMALIES gives each sales series an expected range based on its recent history, and each new reading is compared with that range at a confidence level the team chooses — without training a separate model for every series. A reading outside the expected range flags the series for attention.

Supply chain planning. With AI_FORECAST, a pretrained model can forecast demand across products and stores — in the announcement’s example, a 28-day forecast with a P10–P90 range, refreshed as new sales data arrives. That forecast is the supply-chain planning building block the announcement describes.

Fraud and security. The announcement names fraud and security as one of the six use cases, but the supplied research does not describe the mechanism in detail. As an illustration of how the building blocks could be combined — not as documented behaviour — a bank could compare each payment with an expected spending range based on the card’s recent history and feed out-of-range transactions to its existing fraud systems as one more signal.

Trading. An execution desk needs to divide a large order into smaller trades while managing its effect on the market. On each one-minute bar, PatchTST-FM forecasts trading volume for the next 30 minutes with a P10–P90 range. The announcement’s example uses that forecast to inform the execution schedule for a 100,000-share order.

Semantic search. A reliability engineer investigating an unusual sensor pattern may want to know whether similar behaviour appeared elsewhere in the plant. In a separately built application, TSPulse representations could be indexed to retrieve historical patterns across equipment — then the engineer inspects the associated maintenance records and operating conditions.

What changes for teams building with AI agents

For teams building agent systems, the practical change is that predictions become live context rather than batch reports. Agents that worked from data refreshed on a schedule can now act on forecasts and anomalies that update as events arrive — and in the Streamhouse pattern, agents act on the live context while people approve the exceptions.

The same building blocks apply across the four industries the announcement names — manufacturing, retail, banking and airlines — through the six use cases above: production optimisation, anomaly detection, supply chain planning, fraud and security, trading and semantic search.

The development economics change too. Instead of building a separate model for each of thousands of series — one per product, per store, per card, per sensor — a team starts from one pretrained model. TTM makes this concrete: it is the family’s efficiency model, running many series at low cost on CPU.

Limits and open questions

Several things the announcement does not establish. Forecast quality depends on the series and the task, and no accuracy benchmarks for the models running inside Flink were published with the launch — the claims about which model is strongest come from IBM’s own descriptions, not independent evaluation. Pricing and the eligible AWS regions are not established by the supplied research and need confirmation.

Availability needs checking. The September announcement described Confluent Platform support as planned next, while the November post describes on-premises use via Confluent Platform for Apache Flink — teams planning an on-premises deployment should confirm the current status in Confluent’s documentation. It is also worth confirming which model versions the managed functions actually expose: the September announcement lists all four models, but the functions’ documented model list should be checked before building on a specific one.

Finally, the launch does not provide task-specific validation, so teams should evaluate before relying on the models: how does the forecast perform on their own series, where does a decision need causal explanation rather than a forecast, and what happens when a prediction is wrong and no human reviews it? The launch’s own framing is the honest one: models forecast, detect and optimise; agents act on the live context; people approve the exceptions.