Summary
The repository already has solid MCP primitives for discovery, instantiation, evaluation, forecasting, and code export, but it does not currently ship a runnable example that ties those together into a simple agent-style model-selection loop.
Why this matters
One of the clearest ways to demonstrate the value of sktime-mcp for agentic workflows is a truthful end-to-end example where an LLM-style loop:
- discovers candidate forecasters
- evaluates them on a dataset
- picks a winner based on a metric
- generates a forecast
- exports reproducible code
This is directly aligned with the project's agentic positioning and gives contributors/users a concrete workflow artifact to run and inspect.
Proposed addition
Add a new example script and smoke test covering a minimal model-selection loop on a built-in dataset using the existing MCP tool surface, without introducing any new framework or tool abstraction.
Suggested scope
- new runnable example in
examples/
- smoke test in
tests/
- use current tools only:
list_estimators, describe_estimator, instantiate_estimator, evaluate_estimator, fit_predict, export_code
Summary
The repository already has solid MCP primitives for discovery, instantiation, evaluation, forecasting, and code export, but it does not currently ship a runnable example that ties those together into a simple agent-style model-selection loop.
Why this matters
One of the clearest ways to demonstrate the value of
sktime-mcpfor agentic workflows is a truthful end-to-end example where an LLM-style loop:This is directly aligned with the project's agentic positioning and gives contributors/users a concrete workflow artifact to run and inspect.
Proposed addition
Add a new example script and smoke test covering a minimal model-selection loop on a built-in dataset using the existing MCP tool surface, without introducing any new framework or tool abstraction.
Suggested scope
examples/tests/list_estimators,describe_estimator,instantiate_estimator,evaluate_estimator,fit_predict,export_code