Winsorization, also spelled winsorisation, replaces values beyond a chosen boundary with the boundary itself. At the 1st and 99th percentiles, every observation below the 1st percentile becomes the 1st percentile value and every observation above the 99th becomes the 99th. Nothing is deleted. The extreme observations stay in the sample, they just stop being extreme.
The reason it exists is that a mean and a standard deviation are both fragile. Take a universe of 500 stocks where one posts a price-to-earnings ratio of 1,400 after a collapse in earnings. That single value lifts the universe mean and inflates the standard deviation, which compresses every other stock’s z-score toward zero. The factor still exists in the data and no longer separates anything.
Common bounds
| Method | Typical setting | Effect on a 500-stock universe |
|---|---|---|
| Percentile | 1st and 99th | Clips about 5 stocks at each end |
| Percentile | 5th and 95th | Clips about 25 stocks at each end |
| Standard deviation | ±3σ from the mean | Clips whatever falls outside, count varies |
| Standard deviation | ±2σ | More aggressive, reshapes the distribution |
The 1st and 99th percentile pair is the usual default. Wider bounds preserve more of the real distribution; tighter bounds stabilise the score more but start editing data that is not actually erroneous.
Winsorising versus trimming
Trimming removes the outlier rows entirely. Winsorising keeps them at the boundary. For cross-sectional factor scoring, winsorising is generally preferable because trimming shrinks the universe on exactly the dates when the market is most dislocated, and a stock removed from scoring is a stock the strategy cannot rank at all.
The cost you are accepting
The honest objection to winsorisation is that it hides real information. A price-to-earnings ratio of 1,400 is not a data error. It is the market telling you something specific about that company’s earnings, and clipping it to the 99th percentile turns a distress signal into an ordinary reading. When YESBANK was repricing through 2019 and 2020, its financial ratios were extreme because the situation was extreme.
So winsorise the scoring input, and keep a separate rule for the risk. If a stock’s leverage or valuation is genuinely off the scale, the right response is usually a filter that excludes it from the universe, not a clip that quietly folds it back into the pack.
Check the effect either way. Run the same strategy with winsorisation on and off and compare turnover, the identity of the top-ranked names and the realised drawdown. If the two versions produce materially different portfolios, the outliers were carrying real weight in your signal and you should decide deliberately what to do with them.