Bias fingerprints

The framework

A factor table is usually all an outside reader gets. An allocator reads a manager's IC table, a referee reads a submitted backtest, a desk prices a vendor signal. None of them can run the pipeline behind the numbers, and the literature on look-ahead and survivorship bias is written for someone who can.

A defect is not a scalar amount of inflation. It is a direction in factor space, and the direction is a property of the defect rather than of the study it damages. The framework is a generator that re-runs one study under alternative data conventions, a signature per defect, an inference model that separates signature from unknown true performance, and a validation protocol of four gates that must pass before the diagnostic reports anything.

Read the paper · method · measurements

Two signatures, calibrated

Shift in t-statistic from the point-in-time baseline, for the eight factors that move most under either defect. Both columns come from the same generator, changing one data convention at a time.

The two signatures point in near-independent directions and disagree in sign on value factors, so a table cannot be explained by both. The exact zeros carry more weight than the correlation: a factor built from price and volume never reads a filed figure, so a strong illiquidity result cannot have come from a dating defect whatever else is true of the study.

Conditioning a universe on current index membership behaves in the opposite way. Mean IC falls rather than rises, while amihud_illiquidity flips from t = -1.11 to +2.80 and the low-volatility anomaly inverts. The channel is a forward-looking growth filter, not a survival filter: the biased panel wrongly includes 203 names a day that had not yet been admitted, against 88 it wrongly excludes.

The reference study — signal

The generator the signatures are calibrated on. A 22-factor cross-section on reconstructed point-in-time index membership and SEC XBRL fundamentals keyed on filing date, over 12.5 out-of-sample years.

Factors by information coefficient

Equity curve, net of costs

Portfolio, net of costs

Cost sensitivity — Sharpe at 5bp slippage 0.30, at 10bp 0.24, at 20bp 0.12.

Two results that outlive the study

A decile spread with a t-statistic above 2 looks like a strategy. Regressing that spread on the market decides whether it is one.

The long decile runs a beta of 1.10 against the short decile's 0.99, a tilt positive in 73% of months over a window in which the market compounded at 13.7%. A decile-spread t-statistic reported without its beta decomposition is not a claim about stock selection, here or anywhere.

The second result is dispersion. Across six model seeds with identical economics, IC holds at 0.0164 ± 0.0008 while Sharpe ranges from 0.09 to 0.49. A single backtest Sharpe from a stochastically fitted model is a draw from that distribution, and every performance figure above should be read against its spread.

Universe coverage

Point-in-time index membership against names that could actually be priced and traded. The gap is the residual survivorship limitation: delisted tickers are frequently unavailable from the price source, and it bounds the second signature as well as the study.