Wisdom of Crowds: When Consensus Fails
Table of Contents
The wisdom of crowds in finance comes from combining different pieces of useful information. If five analysts forecast the same quarter’s revenue from different evidence, their errors may offset each other. If all five use a sales feed that double-counts a region, averaging their forecasts will still leave that extra revenue in the estimate.
Before relying on a consensus, ask where the estimates came from, which assumptions they share and what the combined number measures. An average forecast can be useful even when it misses. A traded stock price answers a different question: at what price were buyers and sellers able to trade?
What Makes a Crowd Estimate Useful?
A crowd estimate combines several judgments about the same quantity. The judgments need a common target: next quarter’s reported revenue in U.S. dollars, for example, instead of a mixture of next-quarter sales, next-year sales and sales growth. The arithmetic mean adds the estimates and divides by their number. The median is the middle result after sorting them, or the mean of the two middle results for an even number of estimates.
The mean can benefit when errors above and below the target offset each other. The median is less sensitive to how extreme a single high or low estimate is. Both rules can miss if the estimates tend to err in the same direction. Decide how to combine the forecasts before seeing the answer, and look for independent evidence: five forecasts produced by separate people are still exposed to the same mistake if they start with the same faulty data.
Galton’s ox: The mean and median were different
At a contest in Plymouth, England, entrants guessed what an ox would weigh after slaughter and dressing. Francis Galton analyzed 787 usable entries. His first 1907 report printed a median guess of 1,207 pounds and a dressed weight of 1,198 pounds. His follow-up reported an arithmetic mean of 1,197 pounds.[1] [2]
Economist Kenneth Wallis later checked Galton’s working papers and the organizer’s report. The archival median was 1,208 pounds and the dressed weight was 1,197 pounds. The median was 11 pounds high, and the reported mean matched the actual weight to the reported pound. Both were close. This one contest shows what aggregation can achieve, but it doesn’t tell us which rule will win in a new contest or whether the original guesses were independent.[3]
Worked Example: Independent Evidence vs. a Shared Bad Input
Imagine five analysts forecasting the same quarter’s revenue for a retailer. They submit their forecasts before the earnings release; the company later reports $100 million. In the second scenario, the same five analysts keep their individual judgments but all use a feed that counts a $12 million regional subtotal twice. Assume that error adds $12 million to each estimate.
| Inputs | Forecasts, sorted | Arithmetic mean | Median |
|---|---|---|---|
| Separate evidence | 88, 96, 101, 105, 110 | 100 (error: 0) | 101 (error: +1) |
| Same feed error added | 100, 108, 113, 117, 122 | 112 (error: +12) | 113 (error: +13) |
In the first row, the five forecasts sum to 500, so the mean is 500 ÷ 5 = 100. The median is the third value, 101. The mean happens to get the answer right even though the individual forecasts miss by 6.4 million dollars on average in absolute terms: (12 + 4 + 1 + 5 + 10) ÷ 5. Exact cancellation is a feature of this example.
In the second row, the sum increases by 5 × 12 = 60. The mean is 560 ÷ 5 = 112, and the third value rises from 101 to 113. In both rows, the highest forecast is $22 million above the lowest. The spread offers no clue that the whole group has shifted upward. Switching to the median doesn’t fix the data feed.
How much does adding more forecasts help?
Think about repeating the forecasting exercise on comparable targets. Suppose each forecast error has the same standard deviation, σ, and every pair of errors has the same correlation, ρ. Correlation is about how their errors move together across outcomes; it isn’t a measure of how close the published forecasts look today.
With independent errors, ρ is zero and the standard deviation falls to σ ÷ √n. Take 25 hypothetical forecasters, each with a $10 million error standard deviation and errors that average to zero over repeated outcomes. Independence gives $10 million ÷ √25 = $2 million. If errors instead have pairwise correlation of 0.25, the variance is 100 × (0.25 + 0.75 ÷ 25) = 28 in squared millions of dollars; its square root is about $5.29 million. Adding forecasts under that same correlation assumption brings the standard deviation toward $5 million. The shared part of the uncertainty persists.
A common random error can be positive one quarter and negative the next, increasing uncertainty while averaging to zero over time. A fixed bias, like adding $12 million every time, shifts the average error by $12 million even if the other errors are independent. Before treating an extra forecast as another useful opinion, find out if it adds new evidence.
When Seeing Other Estimates Helps or Hurts
In an information cascade, people act on earlier decisions rather than their own evidence. Imagine two team members vote to delay a launch. Later members see only the votes, not the test results, and also vote to delay even though their own tests favor launching. If they would vote to delay regardless of their test results, the extra votes reveal no new private evidence. Bikhchandani, Hirshleifer and Welch formalized this mechanism in a model of sequential decisions. Sharing one bad data feed is a different route to a similar problem: the forecasts can be biased before anyone sees anyone else’s estimate.[6]
Agreement can grow faster than accuracy. In a study published in 2011, Lorenz and colleagues had 144 ETH Zürich recruits answer six factual estimation questions, five times each. A first private estimate was followed by either no peer feedback, the previous arithmetic mean, or individual estimates and their histories. Estimates clustered closer together with feedback, while the error of the group estimate fell only slightly. The analysis used the mean of log estimates, corresponding to the geometric mean on the original scale; the raw arithmetic mean was one type of feedback the researchers showed participants.[4]
Who sees whose answer matters. In a study published in 2017, Becker, Brackbill and Centola recruited 1,360 people online for tasks such as estimating the calories in a photo of a meal. Participants gave a private first answer, and those assigned to a network saw the average of their immediate peers’ answers before each of two revisions. Across trials where everyone had the same number of connections, the group mean and median became more accurate on average. In networks dominated by a central participant, the crowd could move farther from the answer when that person’s initial estimate pulled it in the wrong direction.[5]
These studies examined factual questions with known answers. They don’t tell us that an investment committee or market will see the same results. They support a useful meeting practice: record a first estimate privately, then ask participants to share the data and reasoning behind revisions. An independent test can change the discussion; a show of hands tells you much less about what the group knows.
Crowd Estimate vs. Market Price
A consensus is calculated from submitted estimates under an aggregation rule. A stock’s last traded price comes from an executed order. People who never submit an order have no direct role in that trade, and the available prices and quantities determine what a new order can execute against. Even a trader with a favorable view may sell to reduce risk. A price can reflect information without giving every view equal weight or revealing why each participant traded.[7] [8]
Cont, Kukanov and Stoikov studied April 2010 trades and quotes for 50 U.S. stocks. Over short intervals, the midpoint between the best bid and ask moved with changes in the balance of orders at those prices. The same order imbalance was associated with a larger midpoint change when fewer shares were available at the best quotes. The midpoint can change without a trade, and the study doesn’t show that a new quote better forecasts earnings or long-run value.[9]
When comparing an analyst’s price target with a traded price, check the horizon and assumptions first. A target for a year from now and a trade today are different quantities, even when both are in dollars per share. A gap alone doesn’t establish that the price is wrong. The efficient market hypothesis addresses how prices incorporate information and why an apparent disagreement doesn’t automatically give an investor a repeatable edge. In prediction markets, a contract price can sometimes be read as an implied probability, but that interpretation depends on the contract and trading conditions. It isn’t a head count of people who think the event will happen.
How to Evaluate a Forecast Consensus
- Define what will settle the forecast. Specify the company, quarter, reporting line, currency, forecast deadline and whether success is measured against the first release or a later revision.
- Trace each estimate to its source. Two reports can still rely on the same management guidance, vendor data or demand assumption. Independent work, such as properly designed channel checks, may add evidence; check what each source actually measures.
- Check the center and test a shared assumption. Compare the mean and median. If the mean is much higher, inspect the high estimates. If both depend on the same sales-growth assumption, rerun the comparison with a plausible lower value.
- Score the method over time. Keep dated forecasts, then compare their signed and absolute errors after release. Test against a simple baseline chosen in advance, such as the prior year’s same-quarter revenue, using comparable targets and lead times.
Common Mistakes
- Counting agreeing analysts instead of counting distinct sources. Multiple models can all repeat the same input error.
- Treating a narrow forecast range as a prediction interval. A range of submitted forecasts is just the distance between those submissions. It doesn’t specify how often the outcome will land there.
- Taking a vivid recent update as confirmation. Many forecasters can overweight the same headline. See availability and recency bias for how a shared focus can distort individual judgment.
Limitations
It is hard to know which reports are independent when their data or assumptions aren’t disclosed. A macroeconomic surprise can affect every analyst at once, and even a strong forecasting record can lose relevance when conditions change. The adaptive markets hypothesis offers a separate way to think about how market relationships can change.
The estimates, their dispersion and a market price can all inform a decision. None tells you whether to trade on its own. Keep track of how the number was produced, what could make it wrong and which evidence would change your view. For the broader role of investor psychology, see the guide to behavioral finance.
Frequently Asked Questions
Sources
- Francis Galton, Vox Populi, Nature, March 7, 1907. Contest and originally published median and weight.
- Francis Galton, The Ballot-Box, Nature, March 28, 1907. Published arithmetic mean.
- Kenneth F. Wallis, Revisiting Francis Galton’s Forecasting Competition, Statistical Science, 2014. Sections 2–3, archival corrections and comparison of estimators.
- Jan Lorenz et al., How Social Influence Can Undermine the Wisdom of Crowd Effect, PNAS, 2011. Experimental design, aggregation and results.
- Joshua Becker, Devon Brackbill and Damon Centola, Network Dynamics of Social Influence in the Wisdom of Crowds, PNAS, 2017. Experimental design, Figure 1 and supplementary methods.
- Sushil Bikhchandani, David Hirshleifer and Ivo Welch, A Theory of Fads, Fashion, Custom, and Cultural Change as Informational Cascades, Journal of Political Economy, 1992. Section II.A.
- Sanford J. Grossman and Joseph E. Stiglitz, On the Impossibility of Informationally Efficient Markets, American Economic Review, 1980. Introduction and Section I.
- SEC Office of Investor Education and Assistance, Understanding Order Types, updated August 18, 2026. Market order and limit order sections.
- Rama Cont, Arseniy Kukanov and Sasha Stoikov, The Price Impact of Order Book Events, authors’ manuscript; published in the Journal of Financial Econometrics, 2014. Sections 3.1–3.2.
