A z-score rescales a raw number into a common unit:
z = (x - mean) / stdev
A stock with a return on equity of 26 percent in a universe averaging 14 percent with a standard deviation of 8 percent has a z-score of (26 - 14) / 8 = 1.5. It sits one and a half standard deviations above the universe average.
In factor work the calculation is cross-sectional, not time-series. On each rebalance date you take the value of one factor across every stock in the universe, compute that date’s mean and standard deviation, and score each stock against its peers on that day. The next rebalance recomputes everything, so a z-score of 1.5 always means the same thing relative to the current cross-section.
The point of standardising is combination. Return on equity in percent and 12-month return as a ratio cannot be averaged as they are. Once both are z-scores, a composite such as 0.5 * z_quality + 0.5 * z_momentum is meaningful, and each factor contributes according to its weight rather than to the accident of its measurement unit.
Reading a z-score
Under a normal distribution:
| Z-score | Percentile | Interpretation |
|---|---|---|
| -2.0 | 2.3 | Bottom of the cross-section |
| -1.0 | 15.9 | Below average |
| 0.0 | 50.0 | At the universe mean |
| +1.0 | 84.1 | Above average |
| +1.65 | 95.0 | Top twentieth |
| +2.0 | 97.7 | Extreme |
Those percentile mappings hold only if the underlying distribution is roughly normal. Financial cross-sections often are not.
The alternatives and when to use them
Rank replaces the value with its position in the ordering, which discards magnitude entirely and is immune to outliers. Percentile does the same on a zero-to-one scale. Min-max rescales linearly into a fixed band and is the most outlier-sensitive of the four. Z-score sits in the middle: it preserves how far apart two stocks are, which rank throws away, at the cost of being distorted by extremes.
Caveats
Both inputs to the formula are themselves estimates, and both are dragged by outliers. One stock with a price-to-earnings ratio of 900 inflates the standard deviation for the entire universe, compressing every other stock’s z-score toward zero and quietly muting the factor. Winsorising before scoring is the usual defence.
Ratios with unstable denominators break the assumption harder. A company posting a small loss produces a negative or absurd P/E, and there is no sensible mean for that column. Use earnings yield rather than P/E, or fall back to rank.
Small universes are the last trap. A z-score computed across 20 stocks has a standard deviation estimated from 20 observations, and it will jump around from one rebalance to the next for no reason connected to the companies.