Estimating Survivorship Bias in Equity Backtests
Summary
The document considers how to estimate survivorship bias in a strategy that buys individual US stocks and sells them after a set holding period. It explains that a robust assessment needs historical records for stocks that were delisted, including firms that defaulted, so the backtest can reflect the securities that were actually available at each point in time.
A proposed shortcut is to model defaults using a probability of default and loss given default, applying an assumed price loss to a share of companies. The answer cautions that a fixed delisting rate cannot capture how defaults and delistings vary over time, citing the unusually high rate during 2008–09. It therefore recommends actual historical delisting data as the reliable basis for estimating the bias. The discussion is brief and gives no quantitative comparison of data sources or alternative models; its conclusion is about estimating historical impact, rather than specifying how to construct a complete survivorship-free dataset.
Key ideas
- A survivorship-aware stock backtest should include securities that were later delisted.
- A fixed default probability and loss assumption may miss time-varying waves of company failures.
- The document identifies actual historical delisting records as the reliable way to estimate survivorship bias.
- The discussion offers no quantitative evaluation of approximate default models.
Tags
Full text
# Simple approach to estimate survivorship bias in backtest # Simple approach to estimate survivorship bias in backtest I am conducting a backtest on a strategy that involves buying individual US stocks and selling them after a specific period. One of the key challenges is addressing survivorship bias. Ideally, the best approach to mitigate this bias would be to include all stocks in the dataset, including those that went bankrupt over time. However, most market data providers I checked do not offer complete datasets for delisted stocks up to the point of their default. This limitation has led me to consider whether there might be simpler alternatives. For instance, one possibility could define a Probability of Default (PD) and Loss Given Default (LGD) model. Using this approach, one could assume that a certain percentage of US companies default each year and model their stock prices as decreasing by a fixed percentage (e.g., X%) in such cases. Are such models commonly used in industry practice, and how accurate are they in addressing survivorship bias? Or is incorporating actual historical data of delisted stocks the only reliable way to estimate the impact of survivorship bias effectively? ## Answer by Chris Taylor (score 3, accepted) https://quant.stackexchange.com/a/81632 Incorporating actual historical data of delisted stocks is the only reliable way to estimate the impact of survivorship bias effectively. The rate and severity of company bankruptcies/delistings varies widely over time (e.g. there was a much higher than usual rate of delistings in 2008/9) so any model which assumes a constant rate of delistings will be horribly inaccurate in some periods.
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.