Signal vs. Noise

Weigh yourself every morning. Tuesday you're up a pound, Thursday down two, Saturday up again. You did not gain and lose a person's worth of flesh in a week. You weighed slightly different amounts of water, food, and yesterday's burrito. The number moved, but nothing real changed — and your job is to notice that before you reorganize your life around Tuesday.

That gap has a name. The signal is the real, lasting thing underneath; the noise is the random jitter on top. Every measurement in real life is signal plus noise stapled together. The skill isn't getting rid of the noise — you can't. The skill is not mistaking the noise for the message.

Why one bad day tells you almost nothing

Try the machine below. Set the sample size to something tiny — say, three data points — and hit "new data" a few times. The average lurches around like a drunk on a boat, throwing up dramatic dips and spikes even though the true value behind it never moved an inch. That panic-inducing crash on the chart is pure randomness. It means nothing. It just looks like it means something, which is worse.

Now drag the slider up. The more data you average together, the calmer the line gets — the wild swings shrink and the average settles close to the truth. This is the whole reason sample size matters: a small sample is mostly luck wearing a trench coat. A big sample lets the random ups and downs cancel out so the real number can show through.

The expensive mistake here has a name too: overfitting. Overfitting is treating the last wobble as the new rule — a theory that explains one weird data point perfectly and predicts the future terribly. Your shop has a slow Monday, so you cancel Mondays. A stock dips for a day, so you sell. You've fit your decision to the noise, and noise never repeats on cue.

What to actually do with a wobble

Run it through the moves. Name the belief — "sales are tanking." Check the base rate — how much does this number normally bounce around even when nothing's wrong? Most things have a quiet, invisible range of random variation, and one bad day usually lives comfortably inside it. Then weigh the evidence: is this outside the usual swing, and has it held for more than a heartbeat?

Then update without collapsing. One bad day is not a trend; it's a data point waiting for company. If the dip repeats across a real stretch of days, the average will tell on it — that's a signal, and you move. If it's a lone spike that's gone by Thursday, you watched noise and stayed calm. Either way you waited for the data to outvote the drama.

So before you react to today's number, ask one thing: would this survive if I looked at it ten more times? The machine above is just that question made visible. A real signal keeps showing up as you collect more data. Noise gets quieter and slinks off. You don't have to be certain — you just have to wait long enough to tell which one you're looking at.