# Confluent AI Forecasting: Check Cloud Preview Limits and On-Prem Setup First

IBM's time-series models can work with streaming SQL, but cloud access, local model servers and data handling need separate checks before a useful forecasting trial.

By Paras Madan | 2026-11-19


IBM's November 18 [time-series walkthrough](https://huggingface.co/blog/ibm-research/time-series-powered-streamhouse) explains how its models connect to Confluent for forecasting and anomaly detection through Flink. The models run on CPUs, with cloud and on-premises deployment paths. That is useful for teams whose operational data already flows through a streaming system.

But those paths are not interchangeable. Before writing SQL, I would establish whether the trial uses a managed cloud feature or model servers your team operates. The answer changes readiness, data handling and who owns a failed request.

## Check the cloud feature's actual status

Confluent's [cloud forecasting documentation](https://docs.confluent.io/cloud/current/ai/builtin-functions/forecast.html) labels `AI_FORECAST` an Early Access Program feature for evaluation and non-production testing. It is not covered by a service-level commitment. The managed function defaults to IBM Granite TTM and does not accept your own model.

Treat that as the boundary for a first project. Build a demonstration or an offline comparison, not an automatic operational decision that assumes production support.

Forecasting is billed through compute-pool CFU usage. A compact model does not make the surrounding service free. Keep the trial bounded by input volume and duration, and record its actual usage before projecting a monthly bill.

The November article is a useful entry point, not proof that every feature it describes became generally available that day. The linked implementation documentation matters more than a broad availability sentence.

## For on-premises use, audit the dependencies

The [self-managed setup guide](https://docs.confluent.io/cp-flink/current/jobs/sql-statements/ai/set-up-ai-functions.html) requires an enabled environment catalog, artifact management, compile workers and a Flink 2.3 compute pool. Models run in separate servers; on amd64, the CPU needs x86-64-v3 support, including AVX2. Model files normally download from Hugging Face at startup, with an offline provisioning option.

Make that a readiness checklist before approving a deployment. Record the intended environment, available compute, model-storage location and owner of the server. A SQL function that exists in documentation is not necessarily registered in your environment.

Start in a test environment with one model and one series. That keeps the initial failure surface small enough to diagnose without involving every downstream application.

## Review where the series values go

The self-managed guide says series values go to the model server, not Confluent. Model downloads do not send statement data to Hugging Face. However, the server uses unauthenticated plain HTTP; the guide recommends restricting access with a Kubernetes NetworkPolicy.

It also warns that environment-default updates replace existing defaults. Running statements retain their previous model URL until redeployed. Preserve other settings and plan the transition rather than assuming an endpoint edit immediately moves every job.

I would have the infrastructure owner review these boundaries before using sensitive operational data. “On premises” describes a location, not a complete access-control policy. Confirm which workloads can reach the service and who can change its configured destination.

## Define a useful forecast before calling it

For the cloud function, timestamps should be evenly spaced, and null values are skipped. The horizon counts future time steps, not a universal unit such as hours. Returned forecasts include means and quantiles, plus model metadata.

Choose a consistent sampling interval and write down what a missing reading means. A missing sensor observation and a genuine zero are different events. Do not silently turn one into the other to make a query run.

For an original example, consider a service's request count sampled every five minutes. A six-step horizon represents half an hour if that sampling remains consistent. Store the forecast with its creation time so you can compare it with observations that arrive later, without accidentally using future information.

## Evaluate the signal before automating the action

Compare the forecast against a simple baseline on historical periods that include normal cycles and disruptions. Keep the rule for judging usefulness fixed before comparing outputs.

For the request-count example, inspect whether a warning arrives early enough to help and how often it sends someone investigating a normal fluctuation. An accurate average forecast can still produce a poor alerting workflow.

My first deployment would produce a reviewable signal, not directly change capacity or spending. Once the team understands the misses, costs and operating limits, it can decide which actions deserve automation. Running a model inside a streaming workflow gets predictions closer to the data; it does not make every prediction a decision.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help AI and data infrastructure teams write clear technical guides that developers can put to work.

[Book a 30-minute call](https://cal.com/paras-madan/30min).
