Estimate-Revision Momentum: A View Against Consensus,
and the Data That Would Test It

In one line: The popular way to trade analyst forecast changes looks too crowded to work, and testing my alternative needs paid data.

Abstract. Trading the sign of analyst earnings-estimate revisions is one of the most published and most productized anomalies in equities. This note argues that the consensus implementation is a decayed, cost-sensitive, price-adjacent trade, and that the residual edge, if one survives, lies in the second moment (the dispersion of revisions and the forced flow that forms around revision clusters), not in the first. It is stated deliberately as a view rather than a result. The estimate-history data required to test it at this engine’s bar is not freely available, and the pipeline’s own attempt to build a revision signal is recorded with an error status for exactly that reason. Knowing the price of the test is part of the view.

1. The consensus signal

Stocks whose consensus earnings estimates are being revised upward tend to outperform over the following weeks; the drift after revisions, like the drift after earnings surprises, is a decades-old and heavily documented result (Givoly and Lakonishok, 1979; Chan, Jegadeesh, and Lakonishok, 1996). It is attractive precisely because it is legible: the signal is a clean, slow-moving number that vendors package and desks trade. That legibility is also the problem.

2. Why I would not trade it the consensus way

Three reasons, each consistent with the pattern the rest of this collection documents. First, crowding: revision momentum is among the 400-plus factors catalogued in the published zoo (Harvey and Liu, 2019), and a signal that legible is a signal that is already in the price by the time it is clean. Second, decay under cost: like the 158 technical alphas and the within-sector books tested elsewhere in this museum, a first-moment revision signal turns over often and bleeds its thin edge to transaction cost once the returns are re-run under a margin of safety. Third, it is price-adjacent: revisions and prices move together, so trading the revision sign is, in large part, trading momentum by another name. That is the crowded territory this engine has repeatedly retired.

3. Where I would look instead

The consensus trades the first moment: the direction of the revision. The second moment is less picked over. Revision dispersion (disagreement across analysts around a revision) and the positioning that forms around revision clusters (who is obligated to trade when a number gaps, and the flow that obligation creates) are mechanisms, not price patterns. That is the same family as the two signals that have actually cleared this engine’s bar, both of which read positioning and flow rather than ranking stocks by past returns. I would test the disagreement and the forced flow, not the sign.

4. The data economics, stated honestly

The pipeline attempted a revision signal; it carries an error status in the hypothesis store because clean estimate-revision history, the kind a serious test requires, is a paid vendor feed, not a free download. The free proxies reconstruct revisions from price, which quietly re-imports the very first-moment crowding the exercise is meant to avoid. So the honest position is to name the data that would settle the question rather than to fake it with a proxy that guarantees a contaminated answer. Refusing the contaminated test is not an absence of work; it is the work.

5. The view, stated plainly

Consensus trades the sign of the revision. I would trade the disagreement around it and the flow forced by it, and I would put no capital on either until it cleared the same margin-of-safety hurdle that has retired every price signal tested here. A view is a hypothesis with a bill attached; this note states both the hypothesis and the bill.