Skip to content

Standard Skill: Statistical Analysis

Statistical Analysis is a Standard Skill: it gets numbers that hold up, not just sums. The agent loads it when you ask whether a change or a difference is real, for a trend, a forecast or seasonality, for outliers, for a correlation, or for more than sums and averages. To pick it yourself, type / and choose statistical-analysis.

  1. It pins down the question and reads the Terms, Rules and Data Model of the Semantic Layer that define the metric and the population, because a statistic on the wrong population is precisely wrong. It says which comparison answers your question.
  2. It gets the data with queries, aggregated in the database where the method allows. If a result was cut at 100,000 rows, it samples or aggregates instead, because a cut result is a biased sample.
  3. It computes every statistic with code in the Sandbox, never by estimating, with the packages the Sandbox has, such as pandas, SciPy and statsmodels.
  4. It checks the result against a matching Benchmark, and whether the conclusion holds in the main segments and in another period. For a check of the whole answer, it may load Validate an Analysis.
  5. It shows and says it, often with a Card: a range or confidence interval rather than a single number, the size of an effect next to its p-value, the sample size when it is small, and what it left out and why. It never claims a cause from a correlation.

When the analysis shows something the Semantic Layer should record, such as a seasonal pattern or a Benchmark worth keeping, it suggests a Context Change Request, but never opens one in this Skill.

This Skill is adapted from the statistical-analysis Skill of Anthropic’s Data plugin, under the Apache License 2.0. Its folder holds the licence and a notice of what we changed; ask the agent to show you /skills/standard/statistical-analysis/NOTICE.md.