Constructing a factor signal

A factor is an idea; a signal is the disciplined process that turns it into numbers you can trade. The construction details often matter as much as the factor itself.

The pipeline

1. Choose a clean metric. For value, perhaps earnings yield. Decide exactly how to compute it and from which data, accounting for reporting lags so you never use a number before it was actually published.
2. Clean the data. Handle missing values, remove or cap extreme outliers (a single bad data point can dominate a ranking), and adjust for corporate actions like splits.
3. Normalise across the universe. Convert raw metrics to comparable scores — commonly by ranking every stock or computing a z-score (how many standard deviations from the average). This makes a price-to-earnings ratio and a momentum score combinable.
4. Neutralise unwanted exposures. A raw value signal might accidentally bet heavily on one sector or on small caps. Quants often neutralise sector and size so the signal expresses the factor and nothing else.
5. Form the portfolio. Typically long the top-ranked names, short the bottom — a long-short book that isolates the factor from the overall market.

Why the plumbing matters

Two researchers testing 'the value factor' can get wildly different results purely from these choices — which metric, how outliers are treated, whether they neutralised sectors. The factor is real, but the implementation determines whether you capture it cleanly or drown it in noise and incidental bets.

The point-in-time rule

The single most important discipline: every input must be point-in-time — known and available at the moment you would have traded. Using a restated earnings figure, or a metric published days later, silently injects future knowledge and is the most common reason a beautiful backtest cannot be reproduced live.