Survivorship Bias in Training Data

Issue: Model Trained Only on Survived Assets; Ignores Assets That Failed or Were Delisted

Frequency: Common

Symptoms

• Historical returns underestimate risk (failed assets missing)
• Overconfidence in asset classes
• Recommendations overestimate achievable returns
• Portfolio volatility higher than model predicts

Root Cause Training data only includes assets that survived to present. Failed assets (bankruptcies, delistings) removed from dataset. Model learns “these assets reliably outperform” but ignores selection bias — assets that didn’t survive were pruned.

Example

Scenario: Index replication model
Training data: S&P 500 constituents from 2020-2024
Missing: 20 companies that were delisted during this period
Model predicts: 10% average annual return
Reality: Actual index return 8% (wealth destroyed in delisted positions)
Impact: Underestimate disaster risk

Key Statistics

• Survivorship bias in returns: 1-3% annually (compounded)
• Volatility underestimation: 15-30% (models don’t see tail risks)

Mitigation Strategies

1. Include Dead Assets: Source historical data that includes delisted companies
2. Adjust for Bias: Apply statistical correction for survivorship bias
3. Stress Testing: Historical VaR should be shocked higher
4. Longer History: Longer time periods include more failures

Metrics

• Return assuming all survived vs. actual (difference = bias)
• Volatility backtest (does realized match predicted?)

Alerts

• Realized volatility >20% higher than predicted → Retrain with dead assets

References

• Survivorship Bias in Investment Returns
• Data Biases in Financial ML