Where does noise collide?

Random, unwanted fluctuation in a physical signal, arising from the thermal motion of charge carriers and setting a floor on what an instrument can resolve. Johnson observed this agitation in conductors in 1928, and Nyquist derived its magnitude the same year, giving the voltage variance now called thermal or Johnson noise. The referent is a physical quantity, a voltage or a current, and noise is the part of it that carries no message; engineers track it through the signal-to-noise ratio, the power of the wanted signal over the power of the noise. Here noise is a fact about the world and the apparatus, not about a person or a market, and it can be lowered by cooling or averaging but never removed.[1][2]

Unit

Fluctuating power or voltage in a physical signal

Recovery means

Averaging and filtering can lift a signal out of the noise, but thermal noise sets a floor that no method fully removes

Measured by

Noise power or voltage spectral density, and the signal-to-noise ratio

Unwanted variability in judgments that should be identical, treated as scatter around the target rather than a shared lean away from it. Kahneman, Sibony, and Sunstein named this in Noise in 2021 and separated it sharply from bias: bias is the average error, the same direction for everyone, while noise is the spread, different judges reaching different verdicts on the same case, or one judge differing from a Monday to a Friday. The referent is a set of human judgments, such as sentences, insurance quotes, or medical reads, and the unit is the variance among them. This is a statistical idea borrowed as a name for human error, but the object is people deciding, not a voltage and not a price, and a fix for bias does nothing about the scatter.[3]

Unit

Variance among judgments of the same case

Recovery means

Not applicable

Measured by

Disagreement among judges on identical cases, gathered in a noise audit

Trading on something other than genuine information, and the meaningless price movement it produces. Black used the word in his 1986 address to mean the opposite of information: noise traders act on signals they wrongly take for news, and their activity is what makes a market liquid while also pushing prices away from value. The referent is trading behaviour and the resulting price, not a physical fluctuation and not the scatter of a judgment. Black argued that noise makes markets possible and makes them imperfect at the same time, so that value can be known only within a factor of two. Here noise is a property of a market, defined against information rather than against bias.[4]

Unit

Price movement not backed by information

Recovery means

Not applicable

Measured by

Deviation of price from fundamental value, and volume attributed to non-informed traders

The part of an outcome that no model can explain, the residual variance left once the systematic signal has been fit. In the bias-variance account set out by Hastie, Tibshirani, and Friedman, the expected error of a prediction splits into squared bias, variance, and an irreducible noise term that comes from randomness in the data itself. The referent is a statistical model and its residuals, and noise is what remains after the structure is taken out, read as error variance. This meaning shares its mathematics with the physics one, since both are variance, but the object is a fitted model rather than a physical instrument, and the noise here is a ceiling on prediction, not a voltage on a wire.[5]

Unit

Residual or irreducible error variance

Recovery means

Not applicable

Measured by

Residual variance after a model is fit, and the irreducible error term in the bias-variance split