Why Two Vendors Disagree on the Same Factor

Ask two commercial data providers for "the momentum factor" on the same stock, on the same day, and it's entirely normal to get two different numbers — sometimes numbers that disagree in sign. This surprises people the first time they encounter it, because factor names sound like they should refer to one well-defined quantity, the way "closing price" does. They don't. A factor name like "value" or "momentum" is a label for a family of closely related but non-identical constructions, and every vendor (and every academic paper) makes its own specific choices within that family.

Take momentum, in principle the simplest factor to define: past return predicts future return. Even here, vendors disagree on the lookback window (twelve months is common, but six and three months both appear), on whether to skip the most recent month before measuring (a standard academic adjustment, to avoid capturing short-term reversal rather than momentum), and on whether to risk-adjust the raw return or use it unadjusted. Value is worse: book-to-price, earnings-to-price, and cash-flow-to-price all get called "value," and even within book-to-price, vendors differ on how stale a company's most recent book value is allowed to be before it's replaced, and on how they treat companies with negative book equity. Add in different universe definitions (which stocks are even eligible), different rebalancing frequencies, and different treatment of missing data, and it becomes clear why a "value factor" from two vendors can correlate only loosely with each other.

This isn't merely academic. A researcher backtesting a strategy that combines a proprietary signal with a vendor's off-the-shelf value factor gets a materially different backtest depending on which vendor's version of value they used — and if they later switch vendors, or the vendor revises its methodology, the strategy's apparent performance can shift for reasons that have nothing to do with the strategy itself. Two published academic papers claiming to study "the same" anomaly, using different data vendors and slightly different construction choices, have in some well-documented cases reached opposite conclusions about whether the anomaly is real.

What this means in practice

The practical response is to never treat a vendor's factor name as a specification — always read the vendor's methodology document for the exact construction rules before combining that factor with anything else, and to test whether a strategy's results are robust to reasonable variations in the construction choices (lookback window, universe, rebalancing) rather than fragile to one specific vendor's exact recipe. When comparing your own research against a published result, matching the paper's precise construction choices, not just the factor's name, is usually the difference between replicating the finding and failing to.