Predict

mcmc predict <spec> <samples> draws posterior-predictive samples from a model you have already fitted. It takes the spec and the posterior samples file from a previous fit, and produces a new samples file of predicted outcomes.

mcmc predict model.toml samples.json -o predictions.json

Flag	Meaning
-o, --out <file>	samples output path
--julia-version <channel>	Julia version/channel to run, overriding the spec
--verbose	show the full raw install/precompile output
--json	print the result as JSON

Like fit, predict runs the model in the backend’s runtime (a Julia subprocess, or CmdStan for Stan), so it is one of the few commands that leave the CLI process.

The [predict] block

Which outcomes to predict is declared in a [predict] table in the spec. targets lists the variables to predict by base name, and what happens to them depends on the backend.

• Turing: each target is blanked to missing, so the sampler draws it from the posterior predictive.
• Stan: the model’s generated quantities block is re-run for every posterior draw, and targets selects which generated variables to keep (for example y_rep); the model must declare them.

You can optionally override the data used for the prediction.

[predict]
targets = ["y"]

[predict.data]
x = [11, 12, 13, 14, 15]

predict.data is merged on top of the spec’s [data], so you can predict at new covariate values without editing the rest of the spec. See the Spec file reference for the full schema.