A z-score measures how many standard deviations a current observation is from its historical mean. It provides a standardized way to assess whether a yield, spread, or auction metric is at an extreme relative to history.
Z-score = (Current value - Mean) / Standard deviation
Interpretation:
On this site, z-scores appear in several contexts:
Z-scores assume roughly normal distributions. For metrics with fat tails (common in financial data), extreme z-scores may be more frequent than a normal distribution implies. The expanding-window approach used in the auction grading system (documented in "Grading US Treasury Auctions") mitigates this by using all available history rather than a fixed window.
A z-score has three inputs. The current value is given. The mean and the standard deviation both come from a chosen window of history, and that choice determines the answer.
A short window makes the score responsive and unstable. Two years of history from a period of rising yields produces a high mean, so a yield that is elevated by any longer standard scores near zero. The measure adapts to the regime, which is useful for spotting short term dislocation and useless for spotting a regime that has gone too far.
A long window makes the score stable and slow. Forty years of Treasury history includes double digit yields, so almost nothing in the recent past scores as extreme. The measure is anchored to a world that no longer applies.
An expanding window is the compromise used for auction grading on this site. It starts from all available history and grows as new auctions arrive, so early observations are scored against thin data and later ones against a deep sample. This avoids the arbitrariness of picking a fixed lookback and accepts that the earliest scores are the least reliable.
No window is correct in general. Stating which one produced a score is part of reporting it honestly.
The formula is simple and its assumptions are not. Four failure modes account for most bad readings.
Fat tails. A normal distribution puts about one observation in twenty beyond two standard deviations. Financial series produce them far more often. A reading of three is genuinely unusual in a normal world and routine in a market one, so the implied probability attached to a z-score overstates how rare it is.
Non-stationarity. The formula assumes the mean and the standard deviation describe a stable process. A yield series that has trended for a decade violates this. During a trend the z-score first flags the move as extreme, then quietly resets as the mean catches up, which produces a signal that fires early and then goes silent.
Serial correlation. Yields move in runs rather than independently. That makes the effective sample size smaller than the observation count, so the estimated standard deviation is too tight and scores are too large.
Autocorrelated inputs to the standard deviation. A period of unusual calm compresses the denominator. The next ordinary move then scores as extreme, not because the value is unusual but because the recent past was quiet.
None of these make the measure useless. They mean a z-score is a ranking device rather than a probability statement.
A percentile rank answers a related question without any distributional assumption. It states what fraction of historical observations sat below the current value.
The two disagree in a useful way. A z-score of +2.5 and a percentile of 97 describe the same observation under a normal distribution. If the actual percentile is 92, the series has fatter tails than normal, and the z-score is overstating how unusual the reading is.
Percentiles have their own weakness. They compress at the extremes, so every observation beyond the historical maximum reports as the 100th percentile regardless of how far beyond it goes. A z-score keeps scaling, which is why the two are best read together.
The practical division is that percentiles answer how rare a reading is, and z-scores answer how far it is. For mean reversion logic the distance matters, because the expected size of the correction scales with it.
There is no universal threshold, and picking one imports the normal assumption. Values beyond plus or minus two are conventionally treated as notable and values beyond three as rare. For financial series both occur more often than the normal distribution implies, so these should be read as rankings rather than probabilities.
No. It means the yield is far from its historical mean over the chosen window. Whether it returns depends on whether the mean is still the right reference. An extreme reading is as consistent with a regime change, in which the mean is about to move, as it is with a stretched market about to revert.
Because auction metrics have no natural lookback period and the number of auctions per tenor is limited. A fixed window would discard usable history for the older tenors and would produce grades that shift whenever an old auction dropped out of the window. An expanding window keeps every observation and makes a given grade comparable over time.