# Signal and noise *Most movement is noise* --- Every number a business produces moves. The question that matters is which kind of movement you're looking at: the ordinary wobble a stable process produces all by itself, or a genuine change in the underlying system. The two demand opposite responses, and most management effort goes into reacting to the first as if it were the second. Statistical process control drew the line decades ago: common-cause variation is the noise a system generates naturally, and chasing it case by case makes things worse; special-cause variation is a real signal, and ignoring it is negligence. The practical tool is embarrassingly simple - a trend with a band around it beats any month-on-month comparison, because the band shows what ordinary looks like. The place I see it most is the monthly P&L: month against last month, month against last year, and a standing invitation to explain variances that are mostly noise - last month was exceptional in its own weird ways, last year in its own good ones, and the credit adjustments and delayed invoicing that distorted both are forgotten by the time the comparison is made. Comparisons against plan are at least made against a number somebody chose deliberately. [[Variance]] runs the whole argument through one month's revenue scare. Two sub-ideas do most of the damage when missed. Small samples first: the number of observations you need scales with the square of the effect you're trying to detect, which is why ten interviews aren't a percentage and an underpowered test kills changes that were actually working - [[Samples]] carries the rule of thumb. And skew: most business quantities can't go below zero but have no ceiling, so the tail pulls the average above the typical case, $\text{mean} \approx 1.6 \times \text{median}$ for moderately uncertain work. That one gap explains why schedules built from honest middle estimates still slip and budgets still overrun - nothing was padded dishonestly, the median was just the wrong number to add up. [[Estimates]] turns it into a working rule, and [[Confidence]] applies the same asymmetry to reading a business case, where the point estimate hides exactly the range that matters. The discipline underneath all of it: before reacting to a number, ask what it was measured against, and whether that baseline was chosen or inherited. ---