How Backtests Can Lie

The common flaws

Each of these can turn a mediocre idea into a beautiful chart.

• Lookahead bias: using information not available at the time
• Survivorship bias: excluding companies that failed or delisted
• Overfitting: tuning rules until the past looks perfect
• Cherry-picked start dates that flatter the result
• Ignoring costs, spreads, taxes and realistic execution
• Testing many variations and reporting only the winner

The standard we hold ourselves to

Any historical result we publish must reconstruct the eligible universe as of each historical date, use only information available then, apply the frozen model version appropriate to that period, include delisted companies where data permits, use split-adjusted pricing, assume realistic costs, and report rolling periods rather than one flattering start date.

Until a reconstruction meets that standard, we publish no historical performance at all. We would rather show nothing than show a convincing chart we cannot defend.

The label that matters

Any such study would be labelled a point-in-time reconstructed backtest — never actual investor performance. Historical simulations are hypothetical, depend on model assumptions and do not represent actual client results. Past performance does not guarantee future results.