Sell-side earnings forecasts cluster tightly around a consensus, and the consensus is wrong in a consistent direction. Neither fact is a mystery, and neither is primarily about analytical ability. Both follow from the incentives the job creates.
The two documented biases
Optimism at long horizons. Forecasts made a year or more ahead are, on average, too high. As the reporting date approaches, estimates are revised downward toward what the company will actually report. The pattern is so regular that the downward revision path through the year is itself a well-documented phenomenon.
Slight pessimism at short horizons. In the final weeks, estimates settle marginally below the eventual result. This is what produces the near-universal “beat” — a large majority of companies exceed consensus in a typical quarter, which is not what an unbiased forecast would produce.
A forecast that is too high early and too low late is not random error. It is a systematic pattern, and systematic patterns come from incentives.
Why the errors run that way
Access. An analyst who publishes a bearish view may find management less available — fewer meetings, less time on calls, less help understanding the business. Access is a genuine input to the work, so protecting it has real professional value. The effect does not require anyone to be threatened; it operates through ordinary reciprocity.
The banking relationship. Research and investment banking are separated by regulation and internal controls, and those controls are real. But a bank whose analysts are consistently negative on an industry is not a natural choice to underwrite that industry’s offerings, and everyone involved understands this without it being discussed.
Career asymmetry. Being wrong alongside everyone else costs an analyst very little. Being wrong alone is career-damaging. The rational response to that payoff structure is to stay close to consensus — which is precisely what produces the clustering. This is the strongest of the three effects and it requires no conflict of interest at all.
Company guidance. Companies steer analysts toward achievable numbers. An analyst whose estimate is far above where management is pointing will hear about it. The consensus a company beats is substantially one it helped construct.
What clustering costs the reader
The visible consequence is that the dispersion of estimates understates genuine uncertainty. Twenty analysts within a narrow range looks like agreement produced by twenty independent analyses. Often it reflects one or two independent analyses and eighteen decisions not to stand out.
This matters because dispersion is widely used as a proxy for uncertainty — in risk models, in position sizing, in assessing how much confidence to place in a forecast. If the dispersion is artificially compressed, the uncertainty is being understated systematically, and most severely for the companies with the heaviest analyst coverage.
It also means a “beat” carries far less information than its treatment suggests. When most companies beat most quarters, beating is the base case. The informative events are the exceptions.
What is still useful
None of this makes sell-side research worthless, and dismissing it entirely is a mistake in the opposite direction.
- The industry knowledge is real. An analyst covering one sector for a decade understands its mechanics, its regulation and its supply chains better than almost any generalist. The descriptive content of a research note is frequently excellent even where the conclusion is compromised.
- Revisions carry more signal than levels. The direction and speed of estimate changes is more informative than where estimates sit, because the biases affect the level more than the change. A sharp downward revision is a genuine update.
- Dispersion changes matter even if the level is compressed. A range that widens suddenly indicates genuine disagreement breaking through the incentive to conform.
- The outliers are worth reading. An analyst well away from consensus has accepted a real career cost to be there. That does not make them right, but it usually means they have a specific argument rather than a marked-up model.
The general pattern
The useful habit here extends well beyond equity research. Whenever a set of forecasts clusters, the first question is whether the forecasters were independent. Weather models, economic projections and analyst estimates all get aggregated into a consensus on the assumption that averaging independent errors cancels them out. That assumption is doing enormous work, and it fails whenever forecasters can see each other’s output and face asymmetric penalties for being wrong alone.
Under those conditions a consensus is not the aggregation of many views. It is one view, plus a great deal of agreement — and it will be revised abruptly rather than gradually when it turns out to be wrong, because everyone updates at once.
That is worth remembering the next time a market moves violently on news that seems too small to justify it. The news may have been small. The distance between the consensus and reality was not.
