How to learn Python for data analysis

This stage makes question formulation a separate skill before pandas syntax dominates. You inspect small CSVs by hand, name what one row means, and sort questions into answerable, partly answerable, and unsupported. It sits first because later failure is usually not knowing which question the columns support.

• rows as records
• primary and unique keys
• column grain
• meaning of missing rows
• question-to-column mapping
• answerable vs partly answerable vs unsupported
• plain-language query plans

You can now: Given a new CSV, you can write 20 questions and mark each as answerable, partly answerable, or unsupported from the columns in under two hours.

Now that questions lead, you learn just enough pandas to answer them: reading files, selecting columns, filtering rows, creating grouped summaries, and joining two tables. It sits here rather than earlier because each verb is tied to a question you already wrote.

• pd.read_csv
• pd.read_parquet
• boolean masks
• .groupby().agg()
• .pivot_table()
• .merge()
• sorting and head
• Series vs DataFrame

You can now: Take a small CSV and replicate a spreadsheet pivot table in pandas with a groupby and an aggregation in an afternoon.

You now meet missing values, wrong dtypes, messy strings, dates stored as text, duplicates, and wide-to-long reshaping. This is the work that consumes most of a real analysis. It comes after the core verbs because you need a reason to clean before you touch the repair methods.

• isna and fillna
• astype
• .str.replace
• pd.to_datetime
• drop_duplicates
• melt and pivot
• categorical dtype

You can now: Take a deliberately messy CSV with mixed dtypes and nulls and produce a clean table with one typed column for each question you asked.

You use visualisations to test the answerable questions from stage one, not to decorate. The stage covers distributions, group comparisons, trends over time, and relationships between numeric columns. It sits here because a plot means little until the table is clean and the question has a shape.

• matplotlib.pyplot
• seaborn.histplot
• sns.boxplot
• sns.scatterplot
• plotly.express
• correlation matrix
• line plots with time index
• small multiples

You can now: Given a cleaned dataset, create three plots and state which one best answers the question, with the trade-off visible in the chart.

You add uncertainty checks to stop treating sample patterns as facts. You learn confidence intervals, t-tests, chi-squared tests, correlation with p-values, and simple linear regression. It sits after exploration because numbers should be checked only after the pattern and its business meaning are clear.

• confidence intervals
• t-test
• chi-squared test
• p-value vs effect size
• correlation vs causation
• statsmodels linear regression
• residuals
• sampling variation

You can now: Run a t-test or chi-squared test on a real comparison and write one sentence about whether the difference matters practically, not just statistically.

The final stage turns notebook experiments into repeatable scripts and projects. You organise code into functions, add data checks, manage dependencies, and record the question each analysis answers. It comes last because reusable structure matters most after you know which steps recur.

• notebook-to-script conversion
• functions with type hints
• pytest data checks
• virtualenv and requirements
• pathlib
• logging
• git commits
• storing processed Parquet files

You can now: Turn a notebook analysis into a script that a colleague can run on a fresh CSV and get the same table and chart.