Abakcus — Book Lists
From the books that show why your intuition about chance is wrong to the textbooks that fix it. 23 books in six steps, four of them free.

Tap a cover to open the book.
23
books, in the order to read them
6
steps, from the first surprise to the history
69
years between the oldest book (1950) and the newest (2019)
4
of them free to read online, legally
Probability is the one part of mathematics where being clever does not protect you. Ask a room of mathematicians how likely it is that two people in a group of twenty-three share a birthday, or whether to change doors in a game show, and a good number of them will get it wrong the first time. Our instincts about chance were built for a world of small numbers and familiar risks, and they fail in ways that are consistent enough to have names.
That is the reason to study the subject, and it is also why the order matters more here than anywhere else. If you start with a textbook, you will learn to push symbols around and still fall for every trap, because nobody has shown you that the traps exist. So this list starts with books that have no exercises at all. Their job is to break your intuition in an entertaining way, so that the formulas, when they come, arrive as the answer to a question you already have.
After those come a first textbook, then problems to practise on, then the theory done properly, then the modern subjects that grew out of it, machine learning, Bayesian modelling and causal inference. The last step is the history, which in this field is unusually good, full of gamblers, astronomers, brewers and codebreakers.
Four of the 23 are legally free and I have marked them. The textbooks assume you have met calculus. If you have not, start with step one here, which needs nothing, and read the best calculus books alongside it. And if you only want one book, take number three if you care about data and number eight if you care about chance.
In a hurry
“I have never studied any of this”The Drunkard's Walk No. 01
“I want to read the news without being fooled”The Art of Statistics No. 03
“I am taking a statistics course right now”Freedman, Pisani and Purves No. 06
“I want to pay nothing”OpenIntro Statistics No. 07
“I want probability done properly”Blitzstein and Hwang No. 08
“I am heading for data science”An Introduction to Statistical Learning No. 17Before the books
Question 1 · Coincidence
How many people do you need in a room before it is more likely than not that two of them share a birthday? Ignore leap years and assume every day is equally likely.
Only 23. With 23 people there are 253 different pairs, and the chance that no pair matches drops just below a half, so a shared birthday has probability about 50.7 percent. With 50 people it is about 97 percent.
Question 2 · Conditional probability
A prize is behind one of three doors. You pick one. The host, who knows where the prize is, opens one of the other two to show it is empty and offers you the chance to switch. Should you?
Yes. Your first pick is right one time in three, and the host's move does not change that. So the other closed door holds the prize two times in three, and switching doubles your chances. I told the story of the argument this caused in Marilyn vos Savant was right all along.
Question 3 · Bayes' rule
One person in a hundred has a certain disease. A test detects 90 percent of real cases, and it also gives a false alarm for 9 percent of healthy people. You test positive. How likely is it that you have the disease?
About 9 percent, not 90. Out of 1,000 people, 10 are ill and 9 of them test positive. Of the 990 healthy people, about 89 also test positive. So only 9 of the 98 positive results are real.
Question 4 · Independence
A fair coin has just come up heads five times running. Is the next toss more likely to be tails, to even things out?
No. The chance is still exactly one half. The coin has no memory, and the long-run balance of heads and tails comes from later tosses swamping the early streak, not from the coin correcting it. Believing otherwise is called the gambler's fallacy.
Step I · Books 01–05
Five books with almost no formulas. Read one before any textbook, so that you know what the subject is trying to protect you from.
No. 01 · Step I
Leonard Mlodinow · 2008
Leonard Mlodinow is a physicist at Caltech, and the title is the mathematician's name for a path made of random steps. His argument is that far more of life than we admit, from the careers of film executives to the reputations of fund managers, is that kind of path, and that we explain it afterwards with stories about talent and judgment.
He makes the case by walking through the history of the subject, and each chapter doubles as a lesson. You meet the gamblers who asked Pascal and Fermat how to divide an unfinished game, Bernoulli's law of large numbers, Bayes's rule for reasoning backwards from evidence, and the bell curve. The three-doors puzzle gets most of a chapter. There is not an exercise in the book, and you finish it knowing most of the ideas a first course covers. If I had to pick one book to start with, it would be this one.
Best forAnyone meeting probability for the first time, at any age.
No. 02 · Step I
Darrell Huff, illustrated by Irving Geis · 1954
Darrell Huff was a magazine journalist, not a statistician, and that is why the book works. In about 140 small pages he goes through the ways numbers are used to mislead. There is the average that is really a mean when a median would be honest, the survey whose sample chose itself, the chart whose vertical axis starts well above zero, and the correlation presented as a cause.
The examples are from the advertising and newspapers of the 1950s and the prices are comically out of date. The tricks are exactly the ones in use today. It has no mathematics beyond percentages, it is illustrated with cartoons, and it has probably sold more copies than any other book about statistics. I also have it on my list of math books you can finish in one sitting.
Best forReaders who want to be harder to fool by the end of the afternoon.
No. 03 · Step I
David Spiegelhalter · 2019
David Spiegelhalter held a chair at Cambridge devoted to the public understanding of risk and has been president of the Royal Statistical Society. He has spent much of his career explaining numbers to journalists, judges and doctors, and this book is what he learned from it. It opens with a real case, the question of whether a murderous English doctor could have been caught earlier from the pattern of his patients' deaths.
Every chapter starts from a question like that. Do busier hospitals have better survival rates? What is the risk of eating bacon? How sure are we about unemployment figures? The methods, among them averages, regression, probability, confidence intervals, hypothesis tests and Bayesian reasoning, are introduced only when a question needs them, with very few formulas. He is frank about how often statistics is done badly, in science as well as in the press. It is the best single account I know of how a statistician thinks.
Best forThe reader who wants one modern book on what statistics is for.
No. 04 · Step I
Charles Wheelan · 2013
Charles Wheelan teaches public policy at Dartmouth and used to write for The Economist, and he says plainly that he found his own statistics classes miserable. The book is his attempt to write the explanation he wanted then. He covers the content of a standard introductory course in the standard order, with the difference that every idea is explained in words and through a story before any symbol appears.
The mean and the median come from a bar where a billionaire walks in. Correlation comes from how a film service guesses what you will like. The central limit theorem, which he calls the LeBron James of statistics, comes from a lost bus full of marathon runners. The formulas are kept in appendices at the ends of the chapters, for readers who want them. It is breezier and more American than Spiegelhalter, and it is the better choice if what you need is help with a class.
Best forStudents who are about to take a statistics course and are dreading it.
No. 05 · Step I
Nate Silver · 2012
Nate Silver became famous in 2008 for forecasting an American presidential election, state by state, more accurately than the professional pundits. Before that he had built a system for projecting the careers of baseball players and had made a living at online poker. His book asks a simple question about all such activities. Why are some fields good at prediction and others terrible?
He goes and looks. Weather forecasters turn out to be honest about their uncertainty and have improved steadily for decades. Earthquake prediction has gone nowhere. Economists and political commentators are confidently wrong at about the rate chance would produce. Out of these case studies comes an argument for thinking in probabilities and updating them as evidence arrives, which is to say for Bayes's rule, explained here with unusual clarity. It is long, and a few chapters can be skipped, but it shows statistics at work in the world better than any textbook.
Best forReaders interested in forecasting: elections, weather, sport, markets.
Step II · Books 06–10
Two for statistics and three for probability. They overlap, so choose one of each according to the voice you like. Three of the five are free.
No. 06 · Step II
David Freedman, Robert Pisani and Roger Purves · 1978, fourth edition 2007
Three teachers at Berkeley wrote this book in the 1970s because they thought the usual first course taught students to calculate things they did not understand. Their book has become the model for how to do it the other way round. It uses hardly any algebra. Ideas are explained in plain sentences, and nearly every one is introduced through a real study.
So you learn about controlled experiments from the trial of the polio vaccine in 1954, about bad samples from the magazine poll that called the 1936 election wrong by a landslide, and about confounding from a famous case in which a university seemed to be discriminating against women applicants and, department by department, was not. The exercises ask you to think and explain, and many have no numerical answer. Statisticians who have taught from every textbook on the market tend to say this is still the best one. Read it slowly and do the review questions.
Best forAnyone who wants to understand statistics and not just pass an exam in it.
No. 07 · Step II · Free edition
David Diez, Mine Çetinkaya-Rundel and Christopher Barr · fourth edition 2019
This is a complete introductory textbook that its authors give away. The PDF is free to download, the source files are open, and a printed copy costs about what a paperback novel does. It was written to replace the two-hundred-dollar books that American colleges assign, and it has been adopted by a great many of them.
It follows the course most universities teach, with data and graphs, then probability, the normal distribution, confidence intervals and hypothesis tests, and regression to finish. The explanations are clear and current, the data sets are real, and there are hundreds of exercises with answers to the odd-numbered ones. The same team has free video lectures and software labs to go with it. It is more conventional than Freedman and has more formulas, which is what you want if you are following a syllabus. I list other books of this kind in free math textbooks.
Best forSelf-learners and students whose course does not assign a book they can afford.
No. 08 · Step II · Free edition
Joseph K. Blitzstein and Jessica Hwang · 2014, second edition 2019
Joe Blitzstein teaches the introductory probability course at Harvard, one of the most popular classes in the university, and this is the book of the course. Its distinguishing habit is what he calls the story proof. Wherever a result can be proved by pages of algebra or by telling a story about what the two sides of the equation are counting, he tells the story, and the result stays in your head.
The coverage is everything a first course needs: counting, conditional probability, random variables, the named distributions and how they relate to one another, expectation, joint distributions, and Markov chains at the end. Common mistakes are flagged as they come up, and each chapter has a large set of problems, many with solutions online. The lectures are free to watch, and the authors host a free version of the whole book. You need calculus. For most readers this is the probability book I would choose.
Best forThe reader who wants one probability textbook and wants it to be a pleasure.
No. 09 · Step II
Dimitri P. Bertsekas and John N. Tsitsiklis · 2002, second edition 2008
The same title as the book above and a different temperament. Bertsekas and Tsitsiklis are professors of electrical engineering at MIT, and this is the text for the probability course that MIT's engineers take. It is compact and exact, with every sentence doing work, and it gets to the point faster than any other introduction I know.
The order is the classical one, and the final chapters go further than most first books, into the Bernoulli and Poisson processes, Markov chains, and both Bayesian and classical inference. The problems are excellent and range from routine to hard, with full solutions for many of them on the publisher's site. The complete course, lectures and recitations included, is free on MIT OpenCourseWare. Choose this over Blitzstein if you prefer concision to conversation, or work through one and use the other for problems.
Best forEngineers and computer scientists who like their mathematics lean.
No. 10 · Step II · Free edition
Charles M. Grinstead and J. Laurie Snell · American Mathematical Society · 1997
A third book with the same name, and the one to pick if you want to pay nothing. Grinstead taught at Swarthmore and Snell at Dartmouth, and they released the full text for anyone to download while the American Mathematical Society sells the printed copy.
Two things set it apart. The authors believed that the best way to get a feeling for chance is to simulate it, so the book is full of small computer experiments: toss a coin ten thousand times and plot the lead, deal hands, run a random walk, and look at what happens before proving anything. And each chapter ends with historical remarks that are essays in their own right, on Pascal and Fermat's letters, on the St Petersburg paradox, on how the central limit theorem was found. It treats discrete and continuous probability side by side and finishes with Markov chains and random walks. If you like to see for yourself, finding π by dropping needles is an experiment in the same spirit.
Best forReaders who like to experiment, and who want the history with the theory.
Step III · Books 11–12
Probability is learned by getting problems wrong and finding out why. Two books of problems with the solutions written out in full.
No. 11 · Step III
Frederick Mosteller · 1965
Frederick Mosteller founded the statistics department at Harvard, and this thin book is what he did for pleasure. There are in fact fifty-six problems. Each takes a few lines to state and a page or two to solve, and the solutions are the point, because he uses them to show how an experienced statistician attacks a question.
The first is about socks in a drawer and the last few need some calculus. Along the way come the birthday problem, a three-cornered duel, a needle dropped on a ruled floor, and a prisoner trying to learn his chances from a guard. You will be confidently wrong several times, and that is the education. It costs very little. I have it on my list of puzzle books as well.
Best forStudents, puzzle lovers and anyone preparing for a quantitative interview.
No. 12 · Step III
David Morin · 2016
David Morin teaches physics at Harvard and is known among students for textbooks that are mostly problems with the solutions written out at length. This one, which he published himself, does the same for elementary probability. It assumes only school algebra for most of its length, with calculus kept to a few marked sections.
Each chapter gives a short, careful account of an idea, among them counting, the basic rules, expectation and variance, the main distributions and the central limit theorem, and then hands over to the problems. There are a great many, and every one has a complete solution that explains the reasoning and often gives a second method. Morin has a talent for finding the place where a beginner will go wrong and stopping to deal with it. Use it beside any of the textbooks in step two, or on its own if those feel too fast.
Best forSelf-learners who want hundreds of worked problems at a gentle level.
Step IV · Books 13–16
Four books for readers who want to know why the methods work. They need calculus throughout, and the last two need some patience as well.
No. 13 · Step IV
Morris H. DeGroot and Mark J. Schervish · 1975, fourth edition 2011
Morris DeGroot founded the statistics department at Carnegie Mellon, and his textbook has been the standard bridge between a first course and graduate work for several generations. Mark Schervish revised it after DeGroot's death. The first half is probability with calculus, and the second half is statistical inference built on it, so you see where every method comes from.
What makes it better than its rivals is balance. Estimation, testing and regression are developed from both the classical and the Bayesian points of view, side by side, without the authors pretending that one of them is the whole truth. The proofs are complete, the pace is steady, and the exercises are numerous and well graded. It is a large and expensive book, and an older edition serves just as well. If you plan to study statistics seriously, this is the place to begin the theory.
Best forStudents who want probability and mathematical statistics in one consistent volume.
No. 14 · Step IV
George Casella and Roger L. Berger · 1990, second edition 2001
This is the book that first-year graduate students in statistics are assigned almost everywhere, and "Casella and Berger" is said as one word. It takes the theory of inference and builds it from the ground: probability, transformations and families of distributions, then sufficiency, point estimation, hypothesis testing, interval estimation and the behaviour of all of these as samples grow large.
It is rigorous without using measure theory, which is what makes it usable straight after a book like DeGroot. The exposition is clear and sometimes dry, and the real substance is in the exercises, which are many and hard. People who have worked through them know mathematical statistics. I would not read it for pleasure, but if you need to understand why a method is optimal, or what can go wrong with it, this is where the answer is written down.
Best forGraduate students, and anyone who needs the standard reference on the theory.
No. 15 · Step IV
William Feller · 1950, third edition 1968
William Feller was a professor at Princeton and one of the people who turned probability into a respectable branch of mathematics. His first volume, which deals only with discrete problems such as coins, dice, cards and random walks, is widely thought to be among the best mathematics textbooks ever written on any subject.
What you notice is the taste. Feller chooses problems whose answers defy expectation and then explains them completely. His chapter on the fluctuations of coin tossing shows that in a long fair game one player is typically ahead nearly the whole time, a fact almost nobody believes until they have read the proof. There are hundreds of examples drawn from physics, genetics and games, and a famous quantity of footnotes. It needs no measure theory and only modest calculus, though it demands attention. A second volume covers the continuous theory and is much harder.
Best forMathematics students, and readers who want to see probability written by a master.
No. 16 · Step IV
E. T. Jaynes · Cambridge University Press · 2003
Edwin Jaynes was a physicist, and he spent forty years on this book without finishing it. It was published after his death, edited by a former student, with some chapters visibly incomplete. It is unlike anything else on this list. Jaynes holds that probability is not about frequencies in repeated experiments. It is the extension of logic to cases where we lack certainty, and its rules are the only consistent way to reason from incomplete information.
He derives those rules from a few requirements of common sense, with a thinking robot as his device, and then rebuilds much of statistics on them. Along the way he attacks the orthodox methods of his day with real ferocity, and whole pages are polemic. Readers either love the book or are exasperated by it. I think it should be read after one has learned the standard theory, not instead of it, and read with a pencil, because when Jaynes is right he is very illuminating.
Best forPhysicists, engineers and anyone drawn to the Bayesian view of the subject.
Step V · Books 17–19
Three books on what statistics has turned into in this century: learning from large data sets, Bayesian modelling by computer, and reasoning about causes.
No. 17 · Step V · Free edition
Gareth James, Daniela Witten, Trevor Hastie and Robert Tibshirani · 2013, second edition 2021
Machine learning, seen from the statistics department, is a set of methods for predicting one thing from many others, and this is the book that explains those methods to people who are not specialists. Two of the authors, Hastie and Tibshirani of Stanford, invented several of the techniques in it. They had already written an advanced treatise on the subject, and this is its deliberately accessible younger sibling.
It starts with linear regression and classification and moves through resampling, model selection, tree-based methods, support vector machines and, in the second edition, deep learning. The mathematics is kept to what is needed to understand why each method behaves as it does, and every chapter ends with a lab that works a real data set in code. There are editions with the labs in R and in Python. The authors give the PDF away. It is the most widely used textbook in its field and deserves to be.
Best forAnyone heading into data science or machine learning.
No. 18 · Step V
Richard McElreath · 2015, second edition 2020
Richard McElreath is an anthropologist who directs a department at a Max Planck Institute, and he wrote this book for working scientists who were taught statistics as a flowchart of tests. His view is that those tests are small machines built for narrow purposes and applied far outside them, and that a researcher is better off learning to build a model for the problem in hand.
The book teaches that from nothing, in the Bayesian way, with code on almost every page. You begin by counting the ways data could have arisen, and by the end you are fitting multilevel models and thinking carefully about what causes what. McElreath writes with unusual wit, and his examples, from the height of foragers to the spread of tool use among islanders, stay in the mind. His lectures for the course are free online. It is demanding, because you have to run the code, and it changes how people work.
Best forScientists who use statistics in their research and suspect they are doing it by rote.
No. 19 · Step V
Judea Pearl and Dana Mackenzie · 2018
For most of the twentieth century statisticians were taught that their subject could describe associations and must stay silent about causes, except in a randomised experiment. Judea Pearl, a computer scientist at UCLA who won the Turing Award, has spent his career arguing that this was an unnecessary surrender. With the science writer Dana Mackenzie he explains his alternative for general readers.
The central tool is a diagram, a set of arrows showing what influences what, together with rules for reading off from it which questions the data can answer. Pearl arranges causal reasoning on a ladder of three rungs, which are seeing, doing and imagining what would have happened otherwise. He uses it to revisit old controversies, above all the decades-long fight over whether smoking causes lung cancer. He is not modest about his own part in the story. It is nonetheless the clearest popular book on the most important recent change in how statistics is understood.
Best forReaders who have been told that correlation is not causation and want to know what is.
Step VI · Books 20–23
Four histories. Probability began at the gaming table and statistics in the observatory, and the people involved were more interesting than those of most sciences.
No. 20 · Step VI
David Salsburg · 2001
The title comes from a summer afternoon in Cambridge in the 1920s. A lady at a tea party claimed she could tell by taste whether the milk had been poured into the cup before the tea or after it, and one of the guests, Ronald Fisher, began to work out how such a claim could be fairly tested. The design of experiments, as a science, grew from that kind of question.
David Salsburg, who worked for years as a statistician in the pharmaceutical industry, tells the story of the people who built modern statistics, with no formulas at all. There is Karl Pearson and his empire in London, the brewer William Gosset, who published under the name Student because his employer forbade it, Fisher and his feuds, and the wave of mathematicians who reshaped the field in mid-century. It is anecdotal and affectionate, and it makes the names in the textbooks into people.
Best forAnyone who has used a t-test and wondered who thought of it.
No. 21 · Step VI
Peter L. Bernstein · 1996
Peter Bernstein was an economist and investment adviser, and his thesis is that the boundary between ancient and modern times is the mastery of risk. Once people believed the future could be measured and managed, and was not simply the whim of the gods, everything from insurance to engineering to capitalism became possible.
He traces that idea through its makers. The story begins with Renaissance gamblers and reaches the exchange of letters in 1654 in which Pascal and Fermat founded the theory of probability. Then come the first life tables, the Bernoulli family, the bell curve, Galton's discovery of regression to the mean, and finally the twentieth-century theories of portfolios and options that run modern markets. Bernstein writes with a storyteller's ease and it was a bestseller. The later chapters lean toward finance. As a history of how humanity learned to think about chance, it has not been bettered.
Best forReaders from business and finance, and lovers of big-picture history.
No. 22 · Step VI
Sharon Bertsch McGrayne · Yale University Press · 2011
In the 1740s an English clergyman named Thomas Bayes worked out a rule for revising a belief in the light of new evidence, and did not publish it. The great French mathematician Laplace found it independently and used it for decades. Then, for most of the twentieth century, the leading statisticians declared the method unscientific, because it requires you to state what you believed before you saw the data.
Sharon McGrayne's book follows what happened in the meantime, which is that people with real problems went on using it quietly. Insurance actuaries used it to set premiums. Alan Turing used it at Bletchley Park against the Enigma cipher. The American navy used it to search for a lost hydrogen bomb and a sunken submarine. Only when computers became fast enough did the method come out into the open. It is told through people, with almost no mathematics. Turing's part in it is covered more fully in his biography.
Best forReaders curious about Bayes's rule and why statisticians fought over it for 200 years.
No. 23 · Step VI
Stephen M. Stigler · Harvard University Press · 1986
Stephen Stigler of the University of Chicago is the leading historian of his subject, and this is the standard work. Its question is a precise one. The mathematics of probability was well developed by 1750. Why did it take another hundred and fifty years before it could be applied to the study of people and societies?
The first half of his answer is set in astronomy, where the problem of combining observations that disagree led Legendre and Gauss to the method of least squares around 1805. The second half follows the long struggle to carry those methods into the social sciences, through Quetelet's "average man" to Galton, Edgeworth and Pearson, who between them invented correlation and regression. Stigler reads the original papers and reproduces their arguments, so some comfort with statistics helps. It is elegantly written. For the wider story it belongs to, see the best history of mathematics books.
Best forSerious readers who want the scholarly history, with the mathematics left in.
After the shelf
Probability leans on calculus from the first textbook onward, and if yours is rusty, the best calculus books are sorted the same way as this list. Later it leans on linear algebra and on the habit of proof, and for the second of those there are the best books for learning mathematical proofs. The whole route, subject by subject, is drawn in how to learn mathematics and in the self-learner's roadmap.
If the puzzles were what you enjoyed, there are many more in the best math puzzle books, and the full story of the three doors is in Marilyn vos Savant was right all along. For the cheapest way to keep going, most of the free math textbooks I recommend have a probability or statistics volume among them.
Keep wandering
A few more pieces in the same spirit — math, design, and slow attention.

The 25 Best Calculus Books to Learn Calculus, in Reading Order
From the 1910 classic that promises any fool can do it to the textbook that makes you prove every line, 25 calculus books sorted by what you need from them.
Open
Marilyn vos Savant Was Right All Along
In 1990, Marilyn vos Savant answered a probability puzzle correctly. Nearly 10,000 people — many of them academics — wrote to tell her she was wrong. She was not.
Open
The 23 Best Math Puzzle Books, from Dudeney to Martin Gardner
Warm-ups, Victorian classics, Martin Gardner, Raymond Smullyan's logic and a last shelf for people who want to suffer. 23 puzzle books in the order I would open them.
Open
Free Math Textbooks from University Mathematicians
125 textbooks that mathematicians give away on their own pages, from high school algebra to graduate analysis, sorted by subject, level and format.
Open
Buffon’s Needle: How to Find π by Dropping Toothpicks
Drop toothpicks on lined paper and count how many land across a line. Out of pure chance comes π, the number of circles, from a game with no circles in it.
Open
The 17 Best Books for Learning Mathematical Proofs
A reading order for anyone moving from computing answers to writing proofs, from the first textbook all the way to Gödel.
Open