Analyst Estimates, Targets and Revisions

What a consensus is, and why it lags

A consensus earnings estimate is the mean of forecasts from analysts covering the company, and its value comes from aggregation rather than from any individual. Two structural forces make it slow. The first is that analysts are paid to be close to consensus: a forecast far from the crowd that turns out right is rewarded modestly, and one that turns out wrong is a career event, so the incentive is to move after news rather than before it. The second is that most forecasts are anchored on company guidance, and guidance is itself a management forecast with its own incentives to under-promise. The predictable consequence is a pattern rather than an edge: estimates are typically revised down through the year and beaten when reported, because guidance is set conservatively and the consensus clusters near it. A company that "beats" by two cents on a lowered bar has told you less than the headline implies. What carries information is the revision in the other direction — an upgrade to full-year estimates from an analyst who had been below consensus, or a change in the trajectory over several quarters, which is why revision momentum is a documented effect while a high target price is not. Dispersion is the third thing to read, and it is usually ignored. When the estimates for next year span a wide range, the analysts are telling you the outcome is genuinely uncertain, and the consensus number conceals it. Wide dispersion also makes a beat or a miss more likely for arithmetic reasons, since the mean of a wide distribution shifts more when one participant changes their mind. A single number cannot carry that information, so look at the range and the number of analysts behind it. One target, decomposed — Target: $120: the headline · On earnings of $6.00: the analyst published the estimate explicitly · So a 20× multiple: the hidden half — and the part anchored to the current price ← · Change the earnings to $5.40: a $108 target on the same multiple — most of the target is the multiple A target is not a prediction of the price in twelve months; it is a statement about fair value on an assumed multiple. When the multiple assumption is anchored to the current price, the target mostly repeats the price back to you.

Using the consensus without trusting it

The productive use is as a benchmark for your own estimates, because it tells you what you have to disagree with. Build your forecast first and compare afterwards; if you look first you will anchor, and a forecast that anchors on the consensus cannot produce an edge over it. Where your number differs materially, the question is whether you know something the analyst does not — a channel check, a segment analysis, a different view of a driver — or whether you have simply been more optimistic for no reason. Second, treat the revisions as the signal and the level as the backdrop. A cluster of upward revisions in the same quarter, especially from analysts whose estimates had been below the average, is the pattern with the better historical record. And keep the analyst’s record in mind: some are systematically early on a sector, and a revision from one of them is worth more than the average, which you can see only by tracking published estimates over time rather than reading the current view. Third, use the target as a stress test of your own valuation. Back out the multiple and the earnings the target implies, and ask what would have to be true. Then ask the same question of the current price. The gap between the two is the market’s implied forecast, and comparing that with your own numbers is a much better use of a consensus number than treating it as a destination (F21 for the machinery of inverting the price into expectations). • Build your own estimate before you look, or you have anchored rather than analysed. • Read revisions and dispersion; discount the level. • Decompose the target into earnings and a multiple, then do the same to the price. • Track an analyst across quarters, not just the current call — reputation is only visible as a history. Regulation separates research from investment banking, but the sell-side still earns from access and trading relationships. That does not make the analysis wrong; it makes the incentive worth knowing when you read a note.

The bias in the consensus has a direction

Consensus estimates are not a neutral average of independent views. They are produced by a small industry with structural reasons to be optimistic, and the optimism shows up in the pattern of revisions rather than in any single number. Three mechanisms drive it. Analysts are paid partly through relationships with the companies they cover, and access depends on asking questions management is willing to answer. Guidance is issued by the company, and most analysts anchor to it because departing from guidance is professionally costly when it goes wrong. And the distribution of outcomes is asymmetric in a specific way: a missed estimate that everyone shares is a market problem, while a missed estimate you alone forecast is a career problem. The consequence is a well-documented tilt: estimates start high and are revised down through the year, so the surprises lean positive late because the bar has been lowered. That is why the revision pattern — and not the level — is the tradable part. A company whose estimates are being raised while the sector consensus is being cut is telling you something about the business; an absolute target that sits 20% above price is telling you what an analyst was paid to say in March. • Estimate levels drift down through the year, so late surprises skew positive • Revisions relative to the sector carry more information than the level • A target that tracks the price is a description, not a forecast • Long-term growth estimates are the least reliable input in the whole model • Coverage dropping to one or two analysts is a signal about the business, not the number The useful question is never whether the consensus is right. It is which part of the consensus you disagree with, and whether that disagreement is large enough to survive costs and the time it takes to be proven.

The numbers inside the number

A consensus figure is a summary of several inputs, and the inputs are visible. The first is **how many analysts contribute**. A consensus built from three estimates is a small sample of opinions, and it moves far more on a single revision than one built from twenty-five. The second is **dispersion**: the spread between the highest and lowest estimate is a direct measure of how much the people doing the work disagree, and it is one of the better proxies for uncertainty available without a model. Wide dispersion around an unchanged mean means the average is hiding a genuine argument, and an event that resolves the disagreement can move the price far more than the mean would suggest. The third input is **staleness**. Consensus datasets mix estimates of different vintages, and an analyst who has not updated since the last quarter is still contributing a number. The consequence is that the consensus is a lagging average, which is why revisions rather than levels carry the information: a rise in the mean driven by several analysts moving in the same direction over weeks is a different signal from one driven by a single stale estimate being refreshed. The practical habit is to look at the count, the dispersion and the date of the most recent changes rather than only at the level. There is a fourth input that matters for the multiple rather than the earnings, and it is easy to miss: the **target price** is usually built from next year’s earnings multiplied by a multiple, and the multiple is frequently just the current one. That means the target is a statement about the earnings forecast wearing a valuation assumption it did not choose. When you decompose a published target, the interesting question is not how far it is from the price but whether the multiple inside it is defensible. If the target is thirty percent above the price and the multiple inside it is unchanged from today’s, the whole call rests on the earnings estimate — and the analysis reduces to whether that estimate is right. Three checks before using any consensus: how many estimates it contains, how wide the dispersion is, and how recent the revisions are. A mean with a small count, a wide spread and old inputs is a number that looks like information and is closer to noise.

Guidance, and the game around it

A company’s own forecast is not a piece of information in the way the accounts are. It is a managed object — set with an eye to what can be beaten — and knowing how it is set and how it is met changes what a headline beat means. What companies guide on is a short list: revenue, earnings per share (usually on an adjusted basis), margins, capital spending, and occasionally a tax rate. What they leave out is as informative: unit volumes, segment detail and the assumptions underneath the range are rarely guided, which is precisely why a quarter can beat the headline while the business behind it does not. Before reading any guidance, find out what is *not* in it. How it is set is a question with a measurable answer, and it is measured per company rather than across the market. Track the last several quarters of guide against actual and the revision that followed. A company that beats by five percent and raises by two has set a lower bar than one that beats by one percent and raises by three, and the market pays for the second. That pattern is the most useful input available for reading the next guide, and it is the reason a beat on its own is close to meaningless. The pattern to fear is a beat accompanied by a cut to the full-year outlook. The headline produces a reaction, and the guidance is the information: earnings above the guide while the forward view comes down means the quarter was strong for reasons the company does not expect to repeat. The mirror pattern, a miss with a raise, is usually the more constructive signal, because it makes the miss unlikely to be a structural problem. Then there is the gap between the consensus and what a stock is priced against. Published estimates are averaged and lag the situation they describe, so consensus is a backwards-looking construction by nature; the buy-side expectation that the price actually reflects is not published at all. The only way to observe it is indirect: compare the size of the consensus beat with the move in the price on the day. A large beat that produces no move tells you the whisper was higher than the published number. Finally, the mechanical reasons a headline beat can be an illusion, and the reconciliation statement is where each one shows up: a lower share count from repurchases, a lower effective tax rate, a helpful currency move, and adjustments that exclude real costs of the business. None of these is illegitimate on its own. Reading a beat without checking the bridge from adjusted to reported earnings is how a company with flat operations gets repriced as a grower. • Guidance is a managed range; the pattern of guide versus actual per company is the input. • A beat with a cut to the outlook is the pattern to fear, and a miss with a raise is the constructive mirror. • Consensus lags; the price reaction reveals the whisper the consensus cannot. • Read the adjusted-to-reported bridge before accepting any beat at face value. A company that withdraws guidance entirely has made a statement about its own visibility. That is a different risk from a lowered number, and it deserves to be treated as one rather than as an absence of information.

Sources

• The distribution of analyst forecast accuracy and the drift of targetsStandard sell-side survey evidence; “Inside the Black Box of Sell-Side Financial Analysts” (Brown, Call, Clement & Sharp, 2015)
• Anchoring and insufficient adjustment in forecastsTversky & Kahneman (1974); experimental evidence on analyst target behaviour
• Estimate revisions and post-earnings driftStandard earnings-surprise literature
• How a target decomposes into earnings and a multipleStandard sell-side methodology