Summarising Findings in a Report

Raw Jupyter notebooks with hundreds of plots and code cells are difficult to share with stakeholders who need to act on findings. A structured EDA summary report distils the most important insights from all your univariate and bivariate analysis into a clear narrative. Good reports separate findings (what the data shows) from implications (what to do about it) and include supporting visualisations only for key claims.

A professional EDA report typically contains six sections: 1) Dataset Overview (shape, source, time period), 2) Data Quality Issues (missing values, outliers, duplicates and their treatment), 3) Key Variable Distributions (most important numeric and categorical columns), 4) Correlations and Associations (highest-magnitude relationships found), 5) Subgroup Insights (group differences that matter for the business question), and 6) Recommended Next Steps (cleaning actions, modelling suggestions).

The overview section should be generated programmatically to ensure it is always accurate. Capture shape, column names and dtypes, date range (if applicable), and the source or version of the dataset. Writing this section in code rather than prose prevents the common error of describing a 10,000-row dataset in a report that was actually generated from a 9,800-row cleaned version.

The data quality section lists every issue found and the decision made for each. Use a table format with columns: Column, Issue, Severity (High/Medium/Low), Action Taken. Compute this table programmatically from the profiling results so it updates automatically when the data changes. Severity should reflect business impact — a missing target variable is High, a missing non-predictive column may be Low.

A summary figure combines several plots into one image that can be embedded in a report. Use plt.subplots(rows, cols) to create a grid and assign each key finding its own panel. Save the figure with savefig() at a DPI appropriate for the output format (150 for screen, 300 for print). Title each panel with a clear finding statement like 'Fare is right-skewed (skew=4.1)' rather than just the column name.

Each key finding should be stated as a business-relevant sentence, not a technical description. Instead of 'The fare column has skewness=4.1', write 'Ticket prices are extremely right-skewed — a small number of first-class passengers paid over 10× the median fare, which may distort revenue analyses that use mean pricing.' This translation from statistical observation to business implication is what makes EDA reports valuable to non-technical stakeholders.

You can write EDA reports as markdown files directly from Python. Build a string containing markdown headings, bullet points, and embedded image links, then write it to a .md file. This approach creates reproducible reports that update automatically when the underlying data changes, making them suitable for integration into data pipeline output artifacts or Git repositories.

Jupyter notebooks can be exported to HTML or PDF using jupyter nbconvert --to html notebook.ipynb. This preserves all plots, code, and markdown cells in a single shareable file that requires no Python installation to view. For a fully automated pipeline, run the notebook programmatically with papermill: papermill input.ipynb output.ipynb -p date '2024-01-01', which parameterises and executes the notebook as a script.

The recommended next steps section is often the most read part of an EDA report. Structure it as a prioritised checklist: P1 (blocking) — issues that must be fixed before any analysis is valid (e.g. misidentified dtypes), P2 (important) — cleaning actions that significantly affect model performance (e.g. outlier treatment), P3 (nice to have) — enrichment ideas and additional data sources to explore. This framing makes it easy for the team to allocate effort appropriately.

ydata-profiling (pip install ydata-profiling) generates a comprehensive HTML report in one call. The report includes overview statistics, variable distributions, correlations, missing value patterns, and duplicate detection — all the sections of a manual EDA report. Use it for rapid exploration at the start of a project, then write a custom summary report to highlight the most relevant findings for your specific business question.

Common EDA report anti-patterns: showing every plot for every column (creates 50-page reports no one reads — select the 5-10 most important), stating facts without interpretation ('age has 177 null values' — so what should we do?), not refreshing the report after cleaning (clean the data, then redo the profiling to confirm issues were resolved), and mixing cleaning code into analysis (keep data cleaning separate from the EDA narrative).

Test your understanding of EDA report writing from this lesson.

In this lesson you learned: structure EDA reports into six sections (overview, quality, distributions, correlations, subgroup insights, next steps), generate multi-panel summary figures and markdown reports programmatically, and translate statistical findings into business-relevant language. Next up we explore advanced indexing techniques — MultiIndex creation and hierarchical selection in Pandas.