Playground

Switch between symmetric, skewed, and bimodal data. Slide the bin count and watch shape emerge — or vanish — while mean and median lines reveal the skew.

Drag the dots along a number line. Pull one far to the right and watch the mean chase it while the median barely moves.

Adjust the spread of a distribution and add outliers. Watch how mean, variance, and standard deviation respond in real time.

Draw repeated samples and see how often the 95% confidence interval actually captures the true mean — and what happens when it misses.

Control slope, noise, and sample size to see how a line fits data and where residuals come from.

Pick a source distribution — uniform, skewed, or bimodal. Collect sample means and watch them converge to a bell curve, regardless of what the source looks like.

Drop an observed z-statistic and watch the tail shade. Then switch to p-curve mode and run 1000 null experiments — the p-values scatter uniformly, with 5% below 0.05 by chance alone.

Slide Cohen's d from 0 to 2 and watch two distributions pull apart. The shaded overlap shrinks while the chance a treated individual beats an untreated one climbs from 50% to 92%.

Slide the sample size on a log scale from 10 to 10,000. Watch the margin of error shrink as 1/√n — then flip to mode B and watch the same tiny effect become 'highly significant' on n alone.

Three financial charts. Pick the one with a real upward trend — then reveal that all three are zero-drift random walks. Dial drift up to feel where 'real' starts.

Two panels. One holds a hidden upward trend; one is pure noise. Identify the trend across rounds — at small n, your accuracy will hover around 50%.

Run K significance tests on pure noise, then watch Bonferroni and Benjamini-Hochberg sweep the false positives away. At K=20 with no correction, expect about one 'hit' — even though nothing is real.

Adjust test accuracy and disease prevalence. See how many of the positive results are actually false alarms — and why rare conditions are hard to screen for.

See why a test that flags 95% of true cases can still be wrong most of the time when the condition is rare — base rate and false-positive rate decide the answer. Switch between disease testing, spam filtering, and fraud detection.

Choose a door. The host reveals a goat. Should you switch? Run a thousand trials and let the numbers settle the argument.

Ice cream sales track drownings perfectly — until you recolor by season and the link vanishes. The DAG below names what you just saw: a third variable driving both.

Four real correlations, each broken by a different mechanism. Click through and watch the taxonomy emerge: confounder, reverse causation, chance, selection bias.

Fit polynomials of increasing degree to noisy data. Watch the training error fall while the test error climbs — and see exactly where the model starts memorizing instead of learning.

See the WWII bomber damage pattern that Abraham Wald used to prove we were reinforcing the wrong parts of the plane.

Simulate 5,000 A/B tests under no real effect. Adjust how often you peek at the dashboard and watch the false positive rate climb from 5% to over 25%.

Slide the power-user fraction from 0 to 30% and watch the mean lifetime climb 20× beyond the median. The 'average user' is the gap between these two numbers.