36 tutorials

Basketball

The NBA's stats API is powerful but picky. Learn to pull it politely, then build standings, ratings and shot charts.

Begin with the ratings-table tutorial to learn the ORtg/DRtg/Net vocabulary the rest of the track speaks, then the home-court and shot-chart builds. The modeling arc — logistic regression, k-NN, cross-validation, ROC curves — uses NBA games as its running example and is the site's most complete from-scratch machine-learning sequence; take it in listed order.

Pull Your First NBA Data with nba_api

Pull NBA standings with nba_api, with the browser headers and retry logic stats.nba.com demands. Includes exactly what to do when the endpoint refuses to answer.

BasketballBeginner~9 min

A current-standings DataFrame from nba_api, with the proper headers baked in.

Filtering and Querying a DataFrame

Use boolean masks and .query() to select the rows matching multiple conditions, then highlight that filtered set on a chart of every NBA team.

BasketballBeginner~5 min

A two-condition filter that finds the contenders, highlighted on a chart.

Grouped Bar Charts: Compare Two Metrics Side by Side

Nudge two bar series apart with a half-width offset to put each NBA team's offense and defense side by side — the chart for comparisons within each group.

BasketballBeginner~5 min

A side-by-side bar chart of offensive and defensive rating for the top NBA teams.

Build a Team Net-Rating Dashboard Table

Combine offensive and defensive ratings into a ranked net-rating table, then style it into a dashboard-quality figure you can drop into a report.

BasketballIntermediate~8 min

Draw an NBA Shot Chart with matplotlib

Draw a regulation half-court from scratch in matplotlib, then plot a player's makes and misses in court coordinates for a real, shareable shot chart.

BasketballIntermediate~10 min

A half-court drawn in matplotlib with a player's makes and misses plotted on it.

NBA Shooting by Zone: Build a Shot-Zone Efficiency Table

Group a season of NBA shots by court zone to see where points really come from: volume, FG%, and points per shot for the rim, mid-range, and three.

BasketballIntermediate~8 min

A table of shot volume and efficiency for every area of the floor.

Draw a League-Wide NBA Shot Heatmap

Plot thousands of NBA shots as a hexbin density map on a court you draw yourself, revealing the modern fingerprint: the rim and the three-point line.

BasketballIntermediate~8 min

A hexbin heatmap of where the entire league takes its shots.

The Efficiency Landscape: Plotting Offensive vs. Defensive Rating

Plot every NBA team's offensive rating against its defensive rating with league-average crosshairs — one scatter separates contenders from rebuilders.

BasketballIntermediate~5 min

A four-quadrant scatter of every NBA team's offense and defense, split by the league-average crosshairs.

Build a Standings Table from a Season of Game Results

Turn a game-by-game results log into a final standings table — one concat plus one groupby collapses a season of scores into wins, losses, and win%.

BasketballIntermediate~8 min

A full W-L standings table aggregated from a game-by-game results log.

A Correlation Heatmap: Which Team Stats Move Together

Compute a correlation matrix with .corr() and render it as an annotated heatmap, showing at a glance which NBA team stats rise and fall together.

BasketballIntermediate~5 min

A correlation matrix of team ratings, drawn as an annotated heatmap.

Crosstabs: Build a Contingency Table to Test Home Advantage

pd.crosstab() counts how two categorical variables co-occur — here, who won against how big the margin was — and normalize='index' turns counts into rates.

BasketballIntermediate~5 min

A home-vs-away win table broken down by margin of victory, charted as home win share.

Bootstrap a Confidence Interval: Is Home-Court Advantage Real or Luck?

NBA home teams won 54.3% of 2023-24 games. Bootstrap the games thousands of times and read a 95% confidence interval that sits entirely above 50%: a real edge.

BasketballIntermediate~9 min

A bootstrap 95% confidence interval on the NBA home win rate, with a histogram of resampled estimates showing the whole interval sitting above 50%.

Build an Elo Rating System from Scratch

Code Elo in about 20 lines: run a real 2023-24 NBA season in date order and watch the ratings discover the standings on their own, plus a home-court edge.

BasketballIntermediate~6 min

Make a Shot-Density Heatmap with hexbin

Turn 25,000 real NBA shots into two hexbin maps — where shots come from, and the make rate by location — exposing the cold mid-range the league abandoned.

BasketballIntermediate~5 min

Two hexbin maps of 25,000 real NBA shots - a density map of where shots come from and an efficiency map of the make rate by location.

Measuring Inequality: Lorenz Curves and the Gini Coefficient

Build a Lorenz curve and Gini coefficient in pure numpy from 25,000 real NBA shots — the top 20% of scorers hold over half the points (Gini ~0.51).

BasketballIntermediate~5 min

A Lorenz curve and Gini coefficient of NBA scoring, showing what share of points the top scorers actually account for.

Is the Difference Real? A Permutation Test

Home teams scored 2.16 more points a game in 2023-24. Shuffle home/away thousands of times to build the null distribution — the edge is real (p < 0.0001).

BasketballIntermediate~5 min

A permutation test on the NBA home-scoring edge: shuffle home/away thousands of times, build the null distribution, and read a p-value off it.

Bayesian Updating: Learn a Win Probability Game by Game

Start from a flat Beta(1,1) prior, feed it 1,231 real NBA games, and watch the posterior sharpen into a tight home-win estimate with a 95% credible interval.

BasketballIntermediate~6 min

A Beta-Binomial belief about the NBA home-win probability updated one real game at a time, with the posterior sharpening from a flat prior into a tight estimate and a 95% credible interval.

The Chi-Square Test from Scratch: Is an NBA Floor Left-Right Symmetric?

Build a 2x2 table from 25,000 real NBA shots and compute the chi-square test by hand in numpy — expected counts, the statistic, and a clean null verdict.

BasketballIntermediate~5 min

A 2x2 contingency table (side of floor x shot outcome) from 25,000 real NBA shots via pd.crosstab, expected counts and the chi-square statistic computed by hand in numpy (no scipy), and the df=1 critical-value verdict - which lands at 0.005 against a 3.841 bar, a textbook fail-to-reject.

Best-of-Seven Math: Why Series Favor Favorites (and by How Much)

Turn a single-game win probability into exact best-of-seven odds in pure numpy, verify it by Monte Carlo, and read the 2024 Finals off the amplification curve.

BasketballIntermediate~11 min

The exact best-of-seven win probability, built from binomial coefficients in numpy and checked twice - against the play-all-seven identity and a 200,000-series seeded Monte Carlo - plus the amplification S-curve and the 2024 Finals worked end to end from real net ratings.

Statistical Power: How Many Games Until You Can Trust a Stat?

NBA home teams won 54.3% of 2023-24 games. Simulate thousands of seasons to see why 41 games almost never detect that edge, and how many an honest test needs.

BasketballIntermediate~10 min

A from-scratch power analysis of the real NBA home edge: a one-sided test whose critical win count is read off a simulated fair-coin null, null-vs-true histograms at 200 games, and a simulated power curve - with the normal-approximation formula agreeing to a hundredth - showing how many games an honest test needs.

Simpson's Paradox: When Every Split Disagrees With the Total

A real Simpson's paradox in 25,000 NBA shots: Markkanen out-shoots Gilgeous-Alexander on twos AND threes yet trails overall — why totals flip, and how to catch it.

BasketballIntermediate~10 min

A genuine aggregation reversal surfaced from 25,000 real NBA shots - Markkanen ahead of Gilgeous-Alexander on both twos and threes yet behind overall - plus the weighted-average arithmetic that causes it and an exhaustive scan finding 53 such reversals among 2,485 high-volume pairs.

Regression to the Mean: Why Hot Starts Cool Off, on Schedule

Split an NBA season into halves and regress one on the other: the slope is 0.77, not 1. Extreme records fade on schedule, and a shrinkage estimate says how far.

BasketballIntermediate~12 min

The 2023-24 NBA season split into 41-game halves and regressed on itself: a 0.77 slope with both extremes fading toward the mean, that slope rebuilt from variance accounting as a 0.82 shrinkage factor that out-forecasts naive persistence, and the curve showing 10-game records regress far harder.

Rest Days and Back-to-Backs: Measuring the Schedule as a Player

Compute rest for 2,462 NBA team-games with one groupby diff: back-to-backs erase home court (54.3% to 47.3%), and rested-vs-tired runs 41.9%.

BasketballIntermediate~7 min

The 2023-24 results log reshaped into 2,462 team-games and each team's rest computed with one groupby diff: back-to-backs are 17.4% of the 2,432 team-games that follow a prior game and erase home court entirely (54.3% baseline falls to 47.3%), a rested-vs-tired matchup runs 41.9% and -2.77 a night across 308 games, and the rest-difference staircase climbs 48.3 to 60.7 - plus the In-Season Tournament final, the row standings drop and schedule math keeps.

Multiple Comparisons: What Survives a 30-Test Screen

Thirty exact binomial tests on one NBA season flag 13 teams. Bonferroni keeps 9, Benjamini-Hochberg 12 - and on the next screen both keep nothing.

BasketballIntermediate~18 min

The exact two-sided binomial and Fisher tests written from scratch, then run as families rather than one at a time: thirty NBA home records flag thirteen teams where an identical league flags 1.21, so Bonferroni keeps 9 and Benjamini-Hochberg keeps 12 - while the same machinery asked whether any single team has a home-court edge of its own flags five and keeps none. Closes with two shooter screens of 191 and 124 tests that land on opposite sides of the same threshold.

The Rise of the Three-Pointer: Two Decades of NBA Shot Selection

Aggregate NBA shot data across twenty seasons to chart basketball's biggest tactical shift — the three-point revolution — and quantify how it changed.

BasketballAdvanced~7 min

A chart of the three-point share of NBA shots, season by season.

Does Shooting More Threes Actually Win Games? Two Decades of Evidence

Across 2004-2024, three-point volume more than doubled while accuracy barely moved — a capstone regression showing an era won on shot value, not better aim.

BasketballAdvanced~10 min

A capstone analysis: a three-panel exhibit pairing the 2004-2024 three-point trend with a value calculation and a league-level regression, landing one honest finding.

K-Means Clustering From Scratch: Find Team Archetypes

Build k-means in pure numpy and watch it sort 30 NBA teams into playing-style archetypes from offensive and defensive rating alone — no labels needed.

BasketballAdvanced~5 min

A pure-numpy k-means that clusters 30 NBA teams into four playing-style archetypes from offensive and defensive rating, with a colored scatter of the groups.

Logistic Regression From Scratch: Predicting a Win, Not a Number

Build logistic regression in pure numpy to predict NBA home wins from the net-rating gap — reaching 67% accuracy against a 54% always-pick-home baseline.

BasketballAdvanced~7 min

A pure-numpy logistic regression that predicts whether the home team wins from the net-rating gap, learned with the same gradient descent from tutorial 70 - reaching 67% accuracy against a 54% baseline.

Train/Test Split and the Confusion Matrix: Grading a Classifier Honestly

Hold out 25% of NBA games and grade the home-win classifier honestly (67.9% on unseen games), then build the confusion matrix, precision, and recall.

BasketballAdvanced~6 min

Take tutorial 71's home-win logistic regression, hold out 25% of games the model never sees, and grade it on those - then go past accuracy to the confusion matrix, precision, and recall.

K-Fold Cross-Validation: A Score You Can Trust

Replace one train/test split with 5-fold cross-validation on the NBA home-win classifier: folds range from 61.8% to 72.4% and average 67.8%. One split can lie.

BasketballAdvanced~5 min

Replace tutorial 72's single train/test split with 5-fold cross-validation - train five times, test on every game exactly once - to get a stable accuracy estimate (67.8%) and, just as important, its spread across folds.

ROC Curves and AUC From Scratch

Sweep the threshold on the NBA home-win classifier to trace a ROC curve, then compute AUC two independent ways — trapezoid area and rank identity — both 0.755.

BasketballAdvanced~6 min

Build a ROC curve and its area (AUC) from scratch in pure numpy for the tutorial-71 home-win classifier: sweep the threshold to trace true-positive vs false-positive rate, then compute AUC two independent ways (trapezoid area and the rank identity) and watch them agree at 0.755.

k-Nearest Neighbors From Scratch

Build k-NN in pure numpy on both teams' net ratings to predict NBA home wins, then sweep k from 1 to 101: k=1 overfits at 55.8%, big k smooths to 66.2%.

BasketballAdvanced~6 min

Build k-nearest neighbors from scratch in pure numpy - no training, just store the games and let the k closest ones vote - on a 2-D feature space (home and away net rating) to predict the NBA home win, then sweep k and compare to the tutorial-71 logistic regression.

A Decision Tree From Scratch

Grow a depth-3 decision tree in pure numpy — Gini impurity, brute-force splits, recursion — on NBA home wins. It scores 64.6%, and every decision is readable.

BasketballAdvanced~5 min

Grow a decision tree from scratch in pure numpy on the NBA home-win problem: measure purity with Gini impurity, brute-force the best (feature, threshold) split, recurse to depth 3, then read the flowchart it learned and compare its accuracy to logistic regression and k-NN.

Principal Component Analysis: Where a Box Score Hides Its Shooting

PCA by hand on 279 NBA players: volume and inside-outside carry 63% of ten box-score rates, and the shooting signal hides in the smallest component.

BasketballAdvanced~28 min

Principal component analysis written by hand - power iteration and deflation, no scikit-learn - on ten per-36 box-score rates for the 279 NBA players with 1,000+ minutes in 2025-26. Two components carry 63% of the variance: volume, which tracks the league's usage rate at r = 0.91, and an inside-outside axis that repeats from season to season better than any single stat. A shuffled-noise test keeps a third. The component ranked last holds 0.18% of the variance and 69.9% of the shooting signal, so a three-component summary erases Stephen Curry's shooting.

Gaussian Mixtures: EM Finds the Court's Zones

EM by hand on 24,947 NBA shot locations: six Gaussians find the court's zones, and 704 shots logged at the rim break the likelihood.

BasketballAdvanced~31 min

A Gaussian mixture fitted by the EM algorithm, written by hand, on the court coordinates of 24,947 shots from 2023-24. Never shown the league's zone labels, six components agree with them on 85.4% of shots, against 67.3% for k-means, because they can draw the corner threes as strips half a foot wide. Then the likelihood shows its flaw: 704 shots logged at exactly the middle of the basket pull a component onto one point, where the likelihood has no maximum, and BIC keeps asking for more components all the way to ten.

Random Forests: A Column of Noise Ranks Third

A random forest in numpy on 923 NBA games: 500 trees vote, and the built-in importance ranks a column of random numbers third of six.

BasketballAdvanced~30 min

The random forest the decision-tree tutorial set as its challenge, written by hand: bootstrap resamples, a random pair of columns at every split, 500 trees voting on 308 held-out 2023-24 NBA games. Trees that average 59.4% alone reach 65.6% together. Then the forest ranks its columns, and its built-in score puts a planted column of random numbers third of six, ahead of both rest columns; Breiman's out-of-bag shuffle test scores it at zero.