Survivorship bias is a composition problem
If today’s successful companies are selected and their past is examined, bankrupt, acquired and delisted participants may disappear. The result answers “how did today’s survivors behave?” rather than “what would an investor have received under the rule at that time?” Closed funds create a similar problem. A record’s disappearance from a current directory does not mean the historical owner had no loss.
Supposons four equal investments produced plus 20%, plus 10%, minus 10% and minus 100%. The average is minus 20%. If the last company disappears from the database and only three survivors remain, the average becomes about plus 6.67%. Almost 27 percentage points came from selection, not skill.
Look-ahead bias is a timing problem
Supposons an annual-profit report is published at the end of March, while the strategy uses its figure to buy at the beginning of January. Even if the profit belongs to the prior year, it may have been unknown in January. The economic period is not the availability date. Data versions that existed at every decision point are required.
Il en va de même to restated accounts, corrected series and a future index composition. If a monthly test selects constituents from a list approved only later, it uses a filter unavailable at the time. Changing a technical date in a table does not recreate missing historical versions; the limitation should remain explicit.
Corporate actions also need a calendar
Splits, dividends, acquisitions and delistings must be handled consistently. An adjusted price is useful for returns, but cannot be treated automatically as the actually observed order price or selection threshold. A departed instrument needs a realistic position-closing result, not silent row deletion.
Vérifiez execution timing separately. A decision made at a closing price after it appears cannot automatically be executed at that same price without additional assumptions. Lag, liquidity and costs change the hypothetical outcome. The smaller the apparent advantage, the more these details matter.
How to organize a data audit
Preserve inclusion and exclusion rules, the complete historical object list, publication dates and sources of changes. Separate “data period,” “publication,” “researcher receipt” and “strategy use.” If a date cannot be reconstructed, narrow the claim instead of inventing a convenient availability time.
Data cleaning does not solve repeated strategy selection. Bailey and co-authors discuss the risk of historical overfitting separately: even a correct database can produce a beautiful random result after many attempts.[108](#source-cv-backtest) Clean data is necessary, not proof of durable profitability.
An audit should report possible biases, corrected test variants and unresolved unknowns. If the result deteriorates after including vanished objects or publication delays, that is a finding, not a reason to restore the former method. This guide teaches research checking; it does not recommend a trading system or promise profitability once the two errors are removed.
Source availability is not the strategy’s right to know
A modern database may be excellent for today’s analysis but unsuitable for reconstructing past decisions. Restated accounts and cleaned classifications are useful to researchers, yet a historical test needs version dates. Without them, describe the work honestly as an approximate experiment, not an exact reconstruction of data available then.
For a survivor sample, check not only which companies disappeared but how ownership ended. An acquisition may provide cash or new securities; a fund closure has its own cash result. Assigning zero return to every disappeared record is also an error. Use the actual event rule or state the uncertainty.
A useful control is to repeat the calculation on a small, manually checked history with several event types. Agreement does not certify the full database, but can expose systematic processing errors before a large test. This matters more than another decimal place in the final return.
Related foundations
Related foundations are the guide on reading financial charts without an illusion of forecasting and the guide on portfolio risk, losses, horizons and joint asset movement.