Nixtla Baseline Review

Overview

Review one concrete Baseline Lab result set. Rank models from observed metrics, show both aggregate and per-series evidence, and keep benchmark conclusions inside the dataset, horizon, and series sample that produced them.

Prerequisites

• Provide a results_*.csv path or a directory containing one.
• Require columns series_id, model, sMAPE, and MASE.
• Treat this repository as a community integration, not an official Nixtla product or an independent reproduction of an entire benchmark.

Instructions

1. Resolve the input with Glob. If several result files match, list their paths and modification context and ask the user to choose; never silently combine separate runs or pick a filename as "latest" without confirmation.

2. Use Read to inspect the header, representative rows, sibling summary, and run_manifest.json or compat_info.json when present. Record the dataset, horizon, series count, models, frequency, and seasonal period that are actually supported by those artifacts.

3. Run the bundled analyzer:

   python3 ${CLAUDE_SKILL_DIR}/scripts/analyze_results.py path/to/results.csv
   

4. Check the analyzer receipt for rejected rows, duplicate series/model pairs, ties, model coverage, and non-finite values. Stop on schema or numeric failures instead of averaging partial data.

5. Rank models by the metric the user requested; otherwise lead with mean and median sMAPE, then report MASE and win counts as supporting evidence. Do not turn a small mean difference into a categorical claim.

6. Inspect series where models disagree or all reported errors are high. State patterns as hypotheses unless timestamps, seasonality, or covariates in the source data actually support them.

7. Give a bounded recommendation for the observed evaluation only. Require out-of-sample validation, cost/latency checks, and operational review before calling any model production-ready.

Metric guardrails

• Lower sMAPE and MASE are better, but neither has a universal "good" threshold.
• Interpret MASE against the scaling denominator and seasonal period used by the run. A model's MASE below 1 indicates lower test MAE than that in-sample naive scale; it does not prove a fixed percentage improvement in every case.
• Do not claim a SeasonalNaive forecast must have MASE exactly 1 on held-out data. See the result and metric contract.

Validation

• Cite the exact CSV and any manifest/summary used.
• Reconcile reported row count with series count multiplied by model coverage.
• Show ties and missing model/series combinations rather than hiding them.
• Preserve units: Baseline Lab writes sMAPE as percentage points.
• Never fabricate TimeGPT, benchmark, or production-performance results.

Output

Return scope and provenance, data-quality receipt, per-model mean/median/std-dev sMAPE, mean/median MASE, series win counts and ties, failure cases, a scoped recommendation, uncertainties, and reproducible next steps.

Error Handling

• No result file: explain the expected filename/schema and point to the Baseline Lab run workflow; do not manufacture sample metrics.
• Several result files: ask the user to select or explicitly approve a comparison across runs.
• Invalid or duplicate rows: report exact row numbers and stop aggregation.
• Missing run metadata: analyze the CSV but label dataset/horizon/frequency unknown unless the filename or sibling artifacts establish them.
• TimeGPT requested without results: state that no comparison is available; never infer hosted-model performance from StatsForecast rows.

Examples

Use prompts that name the evidence and comparison boundary:

Review nixtla_baseline_m4/results_M4_Daily_h7.csv. Rank by median sMAPE,
show ties and missing coverage, and keep conclusions limited to this run.

Compare AutoETS and AutoTheta using both sMAPE and MASE. Identify series where
the metrics disagree and do not use generic accuracy labels.

Resources

• Result schema and metric guardrails
• Analyzer: scripts/analyze_results.py
• Runtime source: 005-plugins/nixtla-baseline-lab/scripts/nixtla_baseline_mcp.py