Skip to content

Guide: Exploratory Data Analysis

A structured way to ask Wizard for a first, rigorous pass over a dataset you haven't worked with before. You don't need to type any of this as a rigid script — the agent's own agentic loop will already investigate and revise as it goes. This guide is useful as a checklist for what a thorough EDA covers, and as a way to steer the agent if its first pass misses something you specifically want checked.

1. Load and inspect

Ask for the shape, columns, and dtypes, plus a look at the first few rows. This is normally the agent's own first move via the deterministic inspect action — no model call needed for it.

2. Data quality

  • Missing values, and their pattern (random vs. concentrated in specific columns or rows).
  • Duplicate rows.
  • Type mismatches — a numeric column stored as text is the single most common surprise in a real dataset.

3. Univariate analysis

Numeric columns:

  • Summary statistics (mean, median, std, min, max).
  • A normality check, not just an eyeballed histogram.
  • Outlier detection (IQR or Z-score) — and a decision about whether an outlier is a data error or a real extreme value, which changes what you do about it.
  • Histograms with a density overlay, and boxplots.

Categorical columns:

  • Value counts.
  • Bar or pie charts for low-cardinality columns; a table for high-cardinality ones, where a chart would just be noise.

4. Bivariate analysis

  • Numeric vs. numeric — a correlation matrix (Pearson and/or Spearman, depending on whether the relationship looks linear), a heatmap, and scatter plots for the pairs that turn out to matter.
  • Numeric vs. categorical — grouped aggregations and boxplots/violin plots split by category.
  • Categorical vs. categorical — a contingency table, and a chi-square test of independence if the question is actually about association, not just description.

5. Multivariate analysis

For datasets with more than a couple of relevant dimensions: pairplots for smaller datasets, or a dimensionality-reduction pass (PCA/t-SNE) for higher-dimensional ones, purely as a way to see structure before deciding what to model.

6. Summary

The point of an EDA isn't the individual charts — it's a synthesized answer to "what's actually going on in this data, and what would I need to be careful about before drawing conclusions from it." Ask explicitly for that summary if the agent's answer stops short of it: data quality issues found, the real trends, and anything that looked like an anomaly worth a second look.