I went looking for the best data science books for Python users this month, and most of them teach a pandas that is now two major versions old. The pandas 3.0 release shipped in January 2026 with Copy-on-Write as the only behaviour and a new default string dtype, and pandas 3.1 followed in October. Nearly every book on the shelf was written for pandas 1.x or 2.x, so the question is not only which book explains data science best. It is also how much of its code needs fixing as you read.
These are the nine I’d buy, from Wes McKinney’s pandas book to a Polars guide, a statistics text and one book with no code in it at all. It’s the hands-on spoke of our machine learning books list, and it skips the general Python titles already in our Python books roundup. I checked every edition, page count and price band against Amazon’s catalog API, then read each book’s companion repo and the pandas release notes in October 2026, so the version notes below come from those files, not from memory.
Picking a first data science book by what you’ll do with it
If you buy one book, make it Python for Data Analysis. It’s the most complete pandas reference in print, the full text is free to read online, and you’ll hit every idea in it during your first month of real work. Pick something else first only when one of these describes you:
- Want the whole toolkit, plotting and scikit-learn included: Python Data Science Handbook
- Already write pandas, want recipes for pandas 2 with Arrow types: Pandas Cookbook, Third Edition
- Starting fresh and pandas isn’t a requirement: Python Polars: The Definitive Guide
- Statistics is the gap: AI-Assisted Statistics for Data Scientists, the 2026 edition
- You manage the people doing this: Data Science for Business
Beginners who’ve never touched data work should start with Bradford Tuckfield’s No Starch book (pick 8) instead of any of these. It’s the shortest book here at 288 pages, it only asks for basic Python and high school math, and each chapter solves one business problem end to end, so you’ll know by chapter four whether you want the heavier books at all.
Checking each book against the pandas you’ll install today
On Python 3.11 or newer, a fresh pip install pandas gives you 3.1 (older Pythons stop at 2.3), so I wanted to know which library version each book was written against. Where a book has an official repo, its pinned requirements tell you more than the blurb does, because that is the version the author actually ran. The table lists what each one targets:
| Book | Edition, year | Pages | Main library | Version in the author’s repo or site | Free to read |
|---|---|---|---|---|---|
| Python for Data Analysis | 3rd, 2022 | 579 | pandas | pandas 2.0.3 pin, Python 3.10 text | Yes, full HTML |
| Python Data Science Handbook | 2nd, 2023 | 588 | NumPy, pandas, Matplotlib, scikit-learn | Online notebooks are the 1st edition (Python 3.5) | 1st edition only |
| Pandas Cookbook | 3rd, 2024 | 404 | pandas, PyArrow | pandas 2.1.4, pyarrow 15, polars 1.3 | No |
| Effective Pandas 2 | 2024 | 594 | pandas | pandas 2 (title) | No |
| Python Polars | 1st, 2025 | 501 | Polars | polars 1.33.0, Python 3.13 | Sample chapter |
| AI-Assisted Statistics for Data Scientists | 3rd, 2026 | 505 | pandas, statsmodels, scikit-learn (and R) | Python 3.10 to 3.14 | Code only |
| Data Science from Scratch | 2nd, 2019 | 403 | Its own modules | Python 3.6 or newer | Code only |
| Dive Into Data Science | 1st, 2023 | 288 | Python (libraries not listed by publisher) | No repo found | Sample chapter |
| Data Science for Business | 1st, 2013 | 413 | None (concepts) | Not applicable | No |
Read the version column before the page count. Not one pandas book here targets pandas 3, and the newest pandas pin is 2.1.4 in the cookbook, so you’ll meet chained assignment and dtype == object checks that behave differently today. That’s not a reason to skip them, because the concepts haven’t moved. It is a reason to read them with the pandas 3 notes further down open in another tab. The two books that dodge the problem do it by design: the Polars guide pins a recent Polars 1.x, and the 2026 statistics edition lists Python 3.14 in its repo badge.
Every ASIN below maps to the edition on its cover, and I linked the newest edition I could match to a paperback listing. Prices are bands from the day I checked, so treat them as a guide and check the live price before you buy.
1. Python for Data Analysis
Wes McKinney started pandas, and his book is still the one I’d keep on the desk when a groupby does something I didn’t expect.
The third edition runs 579 pages and covers NumPy basics, pandas, data cleaning, joins and reshaping, plotting, groupby, time series and a chapter on modeling libraries. McKinney publishes the full text free online, and that open edition was updated for pandas 2.0.0 and Python 3.10 in April 2023. The companion notebooks in the 3rd-edition branch of his repo pin pandas 2.0.3, so only the cookbook below pins anything newer.
Who’s it for? Anyone who’ll write pandas every week, because the reference chapters on indexing and time series are the ones you’ll come back to. Skip the paperback if you’re happy reading in a browser, since the online edition holds nearly the same text, updated for pandas 2.0. The print copy ran about $40 to $45 when I checked, which I’d pay mainly to have something to scribble my own notes in.
2. Python Data Science Handbook
Jake VanderPlas covers the whole stack in one book, which is why it’s the book I’d pick for someone who wants one tour instead of four references.
It moves through IPython, NumPy, pandas, Matplotlib and scikit-learn in that order, so by the end you’ve loaded data, cleaned it, plotted it and fitted a model with the same habits. The second edition is 588 pages from January 2023. One thing tripped me up: the free website and GitHub notebooks everyone links to are the first edition, written and tested with Python 3.5. They’re still useful and they’re MIT-licensed code, but they aren’t the book you’re buying.
Buy it if you want one tour of the full toolkit rather than a deep pandas reference. Pass if Matplotlib and scikit-learn are already familiar, because the pandas chapters are shorter than McKinney’s and you’d be paying for the parts you don’t need. It sat at about $40 to $45 when I checked.
3. Pandas Cookbook, Third Edition
William Ayd and Matthew Harrison rewrote this cookbook for pandas 2 in October 2024, and it’s one of the newest pandas titles from a mainstream publisher.
The repo’s requirements file tells you what the authors care about. Besides pandas 2.1.4 it pins pyarrow 15, Polars 1.3, DuckDB and Ibis, so there’s real coverage of Arrow-backed types and of moving data between pandas and its neighbours. The 404 pages break down into these chapters:
- Foundations, selection and assignment, data types
- I/O, algorithms, visualization, reshaping
- Group by, temporal data, performance tips
- The wider ecosystem, with a foreword by Wes McKinney
I’d hand it to someone who already knows pandas and keeps reaching for apply. The data types chapter, with its PyArrow-backed types, is the one that pays off on pandas 3, where text columns get a new default dtype anyway. Skip it as a first book: recipes assume you know why you want them. Expect about $35 to $40.
4. Effective Pandas 2
Matt Harrison’s book is opinionated in the useful way: it teaches one style of pandas, method chaining, and argues for it on every page.
Effective Pandas 2 is a 594-page paperback from January 2024, self-published by Harrison. Chaining matters more now than when it came out, because pandas 3 makes every indexing result behave as a copy, and code that assigns through a chain of brackets no longer writes back to the original frame. A book that never teaches that habit in the first place ages better than one that does.
Harrison’s own store lists Effective Pandas 3 as a digital book, and its page says all the code was rewritten for pandas 3, with a new section on AI. Amazon also lists a 597-page paperback from September 2026 under the same subtitle, but the listing doesn’t name an edition, so I’ve linked the second edition I could match cover to ASIN. If you want the pandas 3 text today, buy the digital edition from the author. Skip either if you dislike long chains; the whole book is built on them. The paperback ran about $35 to $40.
5. Python Polars: The Definitive Guide
Jeroen Janssens and Thijs Nieuwdorp wrote the book I’d give a newcomer who has no pandas code to maintain.
It’s 501 pages from April 2025 and assumes no pandas experience at all. It covers expressions, the eager and lazy APIs, reading CSV, Parquet, databases and cloud storage, plotting with Altair, hvPlot, plotnine and Great Tables, and the newer corners like Rust plugins and GPU acceleration. The companion repo currently pins Polars 1.33.0 on Python 3.13, so the code tracks the post-1.0 API rather than the renamed functions older blog posts still show.
Skip it if your job runs on pandas, scikit-learn pipelines and colleagues’ notebooks, because you’ll still need pandas fluency first. Stock was listed as scarce and the price at about $55 to $60, level with the statistics book as the priciest here. The free sample chapter on polarsguide.com is a fair test of whether the lazy API clicks for you.
6. AI-Assisted Statistics for Data Scientists
If you’ve seen Practical Statistics for Data Scientists recommended everywhere, this is its third edition under a new name, and it’s the one to buy now.
Same three authors, same subtitle, same crab on the cover. The authors’ old repo for the second edition now carries a notice pointing to the new one and says it won’t be maintained, which settles the naming question. The July 2026 edition grows to 505 pages from 360 and adds chapters on neural networks, deep learning, and generative AI and LLMs after the classic material on sampling, significance testing, regression and classification. Its repo badge lists Python 3.10 to 3.14, and every concept comes with both R and Python code.
Take it if you can write pandas but can’t explain a confidence interval to a product manager. It’s the wrong pick for proofs: it’s written for practitioners, and a dedicated math for machine learning book goes deeper on the theory. The second edition is still listed but showed a ship date rather than stock when I checked. The new one ran about $55 to $60.
7. Data Science from Scratch
Joel Grus builds the tools himself instead of importing them, and that’s the whole point of the book.
The second edition dates from June 2019 and runs 403 pages. Its code needs Python 3.6 or newer and imports from the book’s own scratch package (vectors via scratch.linear_algebra, for example) rather than leaning on the libraries you’d use at work. That makes it the one coding book here that never touches pandas, and the one that explains what a library is doing when you call it.
It’s for the reader who learns by building, or who has to explain gradient descent in an interview. Skip it if you need to be productive this quarter, because hand-rolled code is slower to write and you’ll still learn pandas afterwards. The listing showed a ship date rather than plain stock when I checked, at about $35 to $40.
8. Dive Into Data Science
Bradford Tuckfield’s No Starch book is the gentlest start on this list, and it’s organised by business question rather than by library.
No Starch says you need only basic Python and high school math, and the 11 chapters read like a first year in an analytics team. You’ll get exploratory analysis and forecasting, group comparisons and A/B testing, then binary classification, supervised and unsupervised learning, web scraping, recommendation systems, NLP and a closing look at other languages. Chapter 2 is free as a PDF on the publisher’s page.
At 288 pages from July 2023 it’s short enough to finish, which matters more for a first book than coverage. Skip it if you already work with data: each topic gets one chapter, so it stops where the others start. It ran about $30 to $35, the second cheapest pick.
9. Data Science for Business
There’s no Python in Foster Provost and Tom Fawcett’s book, and I still think it earns a place on a Python list.
The book’s own site says it plainly: “This is not a book of algorithms.” It teaches data-analytic thinking instead, meaning how to frame a problem as classification or regression, how to judge whether a model is worth deploying, and how to talk about expected value with people who sign budgets. Those ideas don’t depend on a library version, which is why a 2013 first edition of 413 pages is still selling.
Read it after your first technical book, or hand it to a manager who keeps asking for “an AI model”. Anyone after code should look elsewhere. Stock was listed as scarce when I checked, at about $25 to $30.
Reading a pandas 2 book on pandas 3
You don’t need to wait for new editions. The concepts in these books still hold, and the code that breaks follows a short list of patterns that the pandas 3.0 release notes spell out. These are the ones you’ll hit first:
- Chained assignment like
df["y"][df.x > 0] = 1no longer changesdf; usedf.loc[df.x > 0, "y"] = 1 SettingWithCopyWarningis gone, because Copy-on-Write is now the only mode- Text columns infer the new
strdtype, sodtype == objectchecks return False - Time zones default to
zoneinfo, notpytz - Offset aliases
M,QandYinresampleanddate_rangeare nowME,QEandYE
Older books hit an earlier wall too, because pandas 2.0 and NumPy 2.0 each removed APIs that pre-2023 book code still calls:
DataFrame.appendandSeries.append: gone in pandas 2.0, usepd.concatiteritems: gone in pandas 2.0, useitemsnp.float_andnp.NaN: gone in NumPy 2.0, usenp.float64andnp.nan
The pandas docs recommend moving to 2.3 first and fixing its deprecation warnings before 3.0, and that’s also the least painful way to follow a book. Make a virtual environment with the version the book’s repo pins, work through it, then upgrade and fix the warnings, which teaches you the migration as a bonus. Our JupyterLab install guide for Ubuntu 26.04 covers the environment side.
Where do you go once pandas is comfortable? The usual next steps are SQL, which our SQL books list covers, pipelines from the data engineering books, or modeling with the deep learning books once scikit-learn feels small.
What I still don’t know
A few things I couldn’t settle from catalogs and repos, and I’d rather say so than guess:
- Whether the September 2026 Effective Pandas paperback is the pandas 3 edition: the page count fits, the listing never says
- What the Data Science Handbook’s second edition changed: the free notebooks are still the first edition
- Which pandas version the printed Python for Data Analysis text names, since the online copy was updated after printing
None of those change the order I’d read them in. Start with Python for Data Analysis or Tuckfield’s beginner book depending on where you are, add the statistics book once you’re making claims from data, and leave Polars for the day a pandas job runs out of memory. If a pandas 3 edition of the cookbook or the handbook appears, I’ll swap it in here.
