Guess the Correlation, Explained
After reading this you will know what Pearson's r measures, why a scatterplot with r near 0.5 looks far rounder than most people expect, and how to train your eye to guess r within about 0.1.
What the game asks you to do
A cloud of dots appears on screen. Somewhere behind that cloud sits a single hidden number, the correlation coefficient r, that controls how tightly the dots hug a straight line. Your job is to look at the cloud and name that number, from -1 to 1. Then the true value appears and you see your error.
Here is the hook. Draw a cloud with true r = 0.5 and show it to ten people. Most will guess something like 0.7 or 0.8. The dots still look loosely lined up to them, so they call it strong. But r = 0.5 is a wide, blurry oval. Half the variation in the vertical direction has nothing to do with the horizontal position. People systematically overestimate the correlation of a cloud that still looks organized, and underestimate how much scatter a "moderate" relationship carries. The game exists to fix that bias in your eye.
What r actually measures
Pearson's r is one number that summarizes how well a straight line describes the link between two variables. It runs from -1 to 1:
- r = 1
- Every point sits exactly on a line that slopes up. Perfect positive relation.
- r = 0
- No straight-line relation at all. Knowing x tells you nothing about y on average.
- r = -1
- Every point sits exactly on a line that slopes down. Perfect negative relation.
Two facts matter and are easy to forget. First, r is about a straight line only. A perfect parabola, dots lying exactly on y = x^2 across a symmetric range, can have r \approx 0 even though the points follow a strict rule. Second, r says nothing about the slope. A gentle upward line and a steep upward line can both have r = 0.9 if the scatter around each is equally tight. Correlation measures tightness, not steepness.
The formula and why it works
Given n paired points (x_i, y_i), the correlation is:
Here \bar{x} and \bar{y} are the means of the two variables. The top is the sum of products of deviations from the mean. When a point sits above average in both x and y, both deviations are positive and their product is positive. When a point sits above average in x but below in y, the product is negative. Add these up and you learn whether the pairs tend to move together (positive sum) or in opposite directions (negative sum).
The bottom is the product of the two spreads. It rescales the top so the result always lands between -1 and 1, no matter the units. Measure height in centimetres or inches, the correlation does not change. That is why r is a pure number with no units attached.
A useful companion is r^2, the fraction of variance in y that a straight line on x accounts for. At r = 0.5 you get r^2 = 0.25: the line explains only a quarter of the vertical spread. That is the arithmetic behind why an r of 0.5 looks so loose.
A worked example you can reproduce
Six points by hand
Take these six pairs, the kind of small set the demo draws around a target r:
| x | y | x - x̄ | y - ȳ | product | (x - x̄)² |
|---|---|---|---|---|---|
| 1 | 2 | -2.5 | -2.5 | 6.25 | 6.25 |
| 2 | 3 | -1.5 | -1.5 | 2.25 | 2.25 |
| 3 | 3 | -0.5 | -1.5 | 0.75 | 0.25 |
| 4 | 6 | 0.5 | 1.5 | 0.75 | 0.25 |
| 5 | 5 | 1.5 | 0.5 | 0.75 | 2.25 |
| 6 | 8 | 2.5 | 3.5 | 8.75 | 6.25 |
- Means: \bar{x} = (1+2+3+4+5+6)/6 = 3.5 and \bar{y} = (2+3+3+6+5+8)/6 = 4.5.
- Sum of products (top):
6.25 + 2.25 + 0.75 + 0.75 + 0.75 + 8.75 = 19.5. - Sum of squared x deviations:
6.25 + 2.25 + 0.25 + 0.25 + 2.25 + 6.25 = 17.5. - Sum of squared y deviations:
6.25 + 2.25 + 2.25 + 2.25 + 0.25 + 12.25 = 25.5. - Bottom: \sqrt{17.5} \times \sqrt{25.5} = 4.183 \times 5.050 = 21.12.
- Divide: r = 19.5 / 21.12 = 0.923.
So these six points carry a strong positive correlation of about 0.923. Notice one point, (3, 3), sits well below the trend, yet the overall r stays high. A single wanderer moves r less than beginners fear when the rest of the cloud is tight.
Train your eye with the widget
The single most instructive thing to watch is the same cloud at different r values. Move r and see the oval fatten or thin. This is what the game is quietly teaching you.
Reading your score and improving
Each round tracks your average absolute error. If you guess 0.7 when the truth is 0.55, your error for that trial is |0.7 - 0.55| = 0.15. Average that over many trials and you get a single skill number.
A useful benchmark: a person who guesses 0 every single time, ignoring the plot entirely, ends up with an average error near 0.5 if the hidden r values are spread evenly across the range, because the mean distance from 0 to a uniform target in [-1, 1] is 0.5. Random guessing across the whole range does worse still, around 0.67. So any honest attempt should already beat 0.5. With practice most people reach an average error near 0.1, which means they can name r to within one decimal most of the time.
Watch the sign of your errors, not just the size. If your guesses are consistently too high when the truth is moderate, you have the classic overconfidence bias. Deliberately talk yourself down: when a cloud looks like a clear 0.8, guess 0.6 and check.
Common mistakes
The number of points changes how a given r looks. With only 15 dots a true r of 0.3 can look like pure noise or like a strong trend by chance. With 500 dots the shape settles to its true oval. Do not read the game's r off a handful of points as if it were solid.
Three traps catch people repeatedly. First, judging steepness instead of tightness: a steep line with fat scatter has lower r than a gentle line with thin scatter, but the steep one looks "more correlated" to a careless eye. Second, letting one outlier hijack your guess. A single far-off point can drag r up or down, but with a big cloud its influence is small, so weigh the bulk. Third, assuming a rounder cloud means no relation. Curvature is invisible to r. If dots follow a clear arc, r can read near zero while the pattern is perfectly real; the game only draws linear clouds, but real data does not promise that.
Related tools
If the correlation game sharpens your feel for how data scatters, several neighbours build the same statistical instincts. The Simpson's Paradox Visualizer shows how a positive correlation inside every group can flip to negative when you pool them, a direct warning about trusting one r. The p-Hacking Simulator shows how a "significant" correlation appears in pure noise if you test enough variables. For the machinery behind confidence in a statistic, try the Bootstrap Resampling Visualizer. To see why averages of anything pile into a bell curve, visit the Central Limit Theorem Demo, and the Galton Board builds that same curve from falling balls. The Law of Large Numbers demo explains why your running score steadies as trials pile up.
Frequently asked questions
Why does r = 0.5 look so much weaker than I expect?
Because r squared, not r, tells you how much of the vertical spread the line explains. At r = 0.5 that is 0.25, so three quarters of the up-and-down variation is unrelated to x. The cloud is a fat oval, and that is correct.
Is a negative correlation weaker than a positive one?
No. Strength is the distance from zero. An r of -0.8 is exactly as strong as +0.8; it just slopes down instead of up. Only the sign differs.
Does high correlation mean one thing causes the other?
No. Correlation measures how two variables move together, not why. A third factor can drive both, or the link can be pure coincidence. r never proves causation.
Can a strong pattern have r near zero?
Yes. If the dots follow a symmetric curve like a parabola or a full circle, the straight-line fit is flat and r reads near 0, even though the pattern is exact. r sees only the linear part.
What average error counts as good?
Guessing 0 every time lands near 0.5. A trained eye reaches about 0.1, meaning you name r to within one decimal most rounds. Below 0.08 is excellent.
Related tools
3D Geometry Board
Points, lines, planes, spheres and polyhedra on an orbiting camera — slice solids with planes, unfold nets, drop the Dandelin spheres into a cone and drag everything.
Bézier Curve Playground
Drag control points and watch De Casteljau’s algorithm construct a Bézier curve step by step.
Central Limit Theorem Demo
Sample from any lopsided distribution and watch the averages pile up into a bell curve.
Conway’s Game of Life
Paint cells and watch Conway’s cellular automaton breed gliders, guns and oscillators.
Dynamic Geometry Board
Construct with points, lines, circles and transformations, then drag anything — the construction stays true, and a checker tells you which relations are theorems.
Fantasy Map Generator
Roll a seed and get a continent: coastlines, mountains, rivers, biomes and named towns, ready for your campaign.