feat: add describe_data_tool for agentic model selection

[kpal002]

Reference Issues/PRs

Ref #386 — Agentic forecaster workflow: missing tools for iterative candidate evaluation.

What does this implement/fix? Explain your changes.

Adds describe_data_tool(dataset, target_col?) — a series fingerprinting tool that returns summary statistics for a named sktime dataset:

• length, n_missing, min, max, mean, std
• trend_slope_per_step — OLS slope of the series vs time index
• candidate_seasonal_period — detected via ACF of the first-differenced series
• frequency — inferred from the pandas DatetimeIndex where available

Why first-differencing matters: raw ACF on a trending series (e.g. airline passengers) is dominated by trend autocorrelation at short lags and returns spurious periods like sp=2. First-differencing removes the trend so seasonal peaks surface correctly (sp=12 for airline data).

Motivation: an LLM agent running a model-selection loop needs to inspect data characteristics before deciding which estimators to try. Without this tool, the agent must call evaluate_estimator or fit_predict blind. describe_data provides the "look at the data first" step that makes iterative agentic selection meaningful. See the linked issue for the full workflow gap analysis.

Changes:

• src/sktime_mcp/tools/describe_data.py — new tool (pure numpy, no new deps)
• src/sktime_mcp/tools/__init__.py — export added
• src/sktime_mcp/server.py — Tool() schema + dispatcher case added

Does your contribution introduce a new dependency? If yes, which one?

No. All helpers use only numpy and pandas, both already core dependencies.

What should a reviewer concentrate their feedback on?

• Dataset loader pattern — I used named dataset strings ("airline", "sunspots", "lynx") to match the convention in evaluate_estimator. If the preferred pattern is data handles from load_data_source, happy to change the interface.
• Seasonality threshold — ACF threshold is 0.2 on the differenced series. Happy to tune or expose as a parameter.
• Scope — read-only fingerprinting tool, no side effects, no handles created.

Any other comments?

This is a draft PR accompanying a proposal for the ESoC 2026 sktime agentic track. The tool is part of a larger agentic forecaster prototype (sktime/sktime#9721) where the same fingerprinting logic has been validated against real sktime datasets including airline passengers.

from sktime_mcp.tools.describe_data import describe_data_tool
result = describe_data_tool("airline")
assert result["success"] is True
assert result["candidate_seasonal_period"] == 12
assert result["frequency"] == "ME"

PR checklist

For all contributions

• I've added myself to the list of contributors.
• Optionally, I've updated sktime's CODEOWNERS to receive notifications about future changes to these files.
• I've added unit tests and made sure they pass locally.