Avoiding Survivorship Bias When Filtering Stocks in Backtests

Summary

The discussion asks whether a stock that has lost money in a daily simulation for two years can be removed without introducing look-ahead bias. The answer cautions against simply dropping persistent underperformers, since real trading universes include firms that later default or delist. Backtests built from only currently listed companies can omit those outcomes and therefore suffer survivorship bias.

The response suggests that prolonged weakness may be relevant information because it can precede delisting, and a live strategy would encounter that path. It does not provide a formal filtering rule, empirical test, or detailed procedure for constructing a point-in-time universe. Any exclusion policy would need to use only information available at each date and account for delisted securities to avoid biasing results.

Key ideas

• Removing stocks based on their observed simulation losses can create selection bias.
• A universe containing only surviving companies may omit defaults and delistings.
• Long underperformance can precede delisting, so the backtest should represent that possibility.
• The discussion offers no specific point-in-time filtering method or supporting empirical analysis.

Tags

# Removing stocks from simulation based on long term out of sample performance


# Removing stocks from simulation based on long term out of sample performance












I have performed a simulation on a stock universe and have found some stocks that out of sample have never performed (every day they always lose money in the simulation). I don't want to introduce any look ahead bias but these stocks have consistently lost money for 2 years (I simulate once a day). Can I remove them from the simulation?

Thanks

## Answer by TomDecimus (score 2, accepted)

https://quant.stackexchange.com/a/42904

Dont remove them, this is a common topic. For example, when you download a set of data, most of the time, you only get companies that did not defaulted/delisted which create a strong bias. Common sense would say, ok if the stock is underperforming for 2 years, it might get delisted and the model should capture this information. Which again, happens in the real world, specially after many months with negative results.

Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.