Sloan Management Review Reprint Series

MANAGING OVERCONFIDENCE

J. EDWARD RUSSO

PAULJ.H.SCHOEMAKER

Winter 1992

Volume33

Number 2

– — — ————-'——� ——-

Managing Overconfidence

1. Edward Russo • Paul 1. H. Schoemaker

f:OOD DECISION MAKING REQUIRES MORE THAN KNOWLEDGE OF FACTS, ';:/ CONCEPTS, AND RELATIONSHIPS. IT ALSO REQUIRES METAKNOWLEDGE­

an understanding of the limits of our knowledge. Unfortunately, we tend

to have a deeply rooted overconfidence in our beliefs and judgments.

Because metaknowledge is not recognized or rewarded in practice, nor in­

stilled during formal education, overconfidence has remained a hidden

flaw in managerial decision making. This paper examines the costs, causes,

and remedies for overconfidence. It also acknowledges that, although

overconfidence distorts decision making, it can serve a purpose during de­cision implementation. 00

]. Edward Russo is Proftssor of Mark�ting

a n d B�havioral Scimu at th� joh n s o n

Graduau School o f Manag�mmt, Corn�//

Univmity. Paul]. H. Scho�k" is Associ­

au Proftssor of Straugy at th� Graduau

School of Businm, th� Univn1ity of Chicago.

They r�cmtly publish�d a book on managm­

al decision making. Decision Traps (Simon

and Schust", 1990.)

To know that we know what we know and that we do not know what we do not know, that is true knowledge. -Confucius

Philosophers and writers have long tried to raise awareness about the difficulty of balancing confi­dence with realism, y et the consequences of un­

supportable confidence continue to plague businesses. Managers deal in opinions – they are bombarded with proposals, estimates, and predictions from people who sincerely believe them. But experience tells managers to suspect the certainty with which these beliefs are stated. For instance: • A leading U.S. manufacturer, planning production ca­pacity for a new factory, solicited a projected range of sales from its marketing staff. The range turned out to be much too narrow and, consequently, the factory could not adjust to unexpected demand. • A loan officer at a major commercial bank felt that his colleagues did not understand their changing competi­tion as well as they thought they did and were refusing to notice signs of coming trouble. • In the early 1970s, Royal Dutch/Shell grew concerned that its y oung geologists too confidently predicted the presence of oil or gas, costing the company millions of dry-well dollars.

SLOAN MANAGEMENT REVIEW/WINTER 1992

• The sales head for Index Technology, a new software venture, repeatedly received unrealistic sales predictions, not only on amounts but also on how soon contracts would be signed.

Managers know that some opinions they receive from colleagues and subordinates will be accurate and others inaccurate, even when they are all sincerely held and persuasively argued. Moreover, given any strongly held opinion, one seldom has to look far to find an op­posing view that is held no less firmly. We do not even have to favor a position now to reserve the right to hold a foture position. One of us attended a faculty meeting at which a senior faculty member had been notably silent during a heated debate. W hen asked for his posi­tion, he replied, "I feel strongly about this; I just haven't made my mind up which way."

People are often unjustifiably certain of their beliefs. As a case in point, the manufacturer cited above accepted the staff's confidently bracketed sales projections of twenty­three to thirty-five units per day and designed its highly automated factory to take advantage of that narrow range. Then, because of a worldwide recession, sales dropped well below twenty-three units per day. The plant was forced to operate far below its breakeven point and piled up enormous losses. Instead of being the best of the com­pany's production facilities, it became the biggest loser.

Russo & SCHOEMAKER 7

Confidence Quiz

For each ot the following questions. provide a low and a high estimate such that you are 90

ranges for questions like: "What was the total dollar value of new com­mercial loans made by [Competitor X] last year?" and "What is the total number of commercial loan officers at the sixth- through tenth-ranking banks in [our city]?"

1 percent certain the correct answer will fall within these limits. You should aim to have 90 , percent hits and 10 percent misses. The correct answers are provided at the end of this arti­cle so that you can compute how close you come to the ideal level of one miss in ten.

90% Confidence Range ,

1. How many patents did the U.S. Patent and Trademark Office issue in 1990?

2. How many of Fortune's 1990 "Giobal500," the world's biggest industrial corporations (in sales). were Japanese?

3. How many passenger arrivals and departures were there at Chicago's O'Hare airport in 1989?

4. What was the total audited worldwide daily circulation of the Wall Street Journal during the first half of 1990?

5. How many master's degrees in business or management were conferred in the United States in 1987?

6. How many passenger deaths occurred worldwide in sche­duled commercial airliner accidents in the 1980s?

7. What is the shortest navigable distance (in statute miles) between New York City and Istanbul?

8. What was General Motors' total worldwide factory sales of cars and trucks (in units) in the 1980s?

9. How many German automobiles were sold in Japan in 19897 1 10. What was the total U.S. merchandise trade deficit with i Japan (in billions) in the 1980s?

lower upper We believe that, like the loan offi­

cer's reluctant boss, you will be sur­prised at how poorly you do on the test below. For each of the ten quanti­ties in the quiz, simply give a low guess and a high guess such that you are 90 percent sure the true value will lie between them. Try it before read­ing further, particularly if you are apt to be confident of your ability to pre­dict accurately.

Metaknowledge

l�——–�——�—-�——–�——– — – – – – ——–The confidence quiz measures some­thing called metaknowledge: an appre­ciation of what we do know and what

A Test of Confidence

How should managers deal with the often unreliable opinions they receive? The answer lies in recognizing that most people's beliefs are distorted by deep-seated overconfidence. Once we understand its nature and causes, we can better devise plans for controlling it. The first step is to document and measure the problem's seventy.

Recall the loan officer who believed that his col­leagues were overconfident about their competitors. He went to his boss with this concern and proposed mea­suring the degree of confidence his colleagues had in their knowledge about the bank's competitors. The boss insisted there was nothing to worry about: "No one is more realistic than a banker." Despite this overconfident answer, the boss agreed to the test – but only he would take it. To his surprise, he failed miserably; to his credit, he then asked all eleven other loan officers to take the same test. Every one of them flunked.

The loan officer's test asked for both best estimates and ranges of confidence around those estimates. The "Confidence Quiz" shown here is just such a test, one that involves general business rather than company­specific questions. The bank test included confidence

8 RUSSO & SCHOEMAKER

we do not know. Normally, we define knowledge as consisting of all the facts, concepts, relationships, theo­ries, and so on that we have accumulated over time. Metaknowledge concerns a higher level of expertise: un­derstanding the nature, scope, and limits of our basic, or primary knowledge. Metaknowledge includes the uncer­tainty of our estimates and predictions, and the ambigu­ity inherent in our premises and world views.1

At times, metaknowledge is more important than primary knowledge. For example, knowing when to see a lawyer or a doctor (metaknowledge) is more impor­tant than how much we know about law or medicine (primary knowledge). We draw on our metaknowledge when we conclude that we have enough information and are ready to make a decision now. If we think we are ready to decide when we are not, we may make cost­ly mistakes. Only when we appreciate the limits of our primary knowledge can we sensibly ask for more or bet­ter information.

E xamining confidence ranges, one of several ways re­searchers study metaknowledge, is a practical means of assessing personai uncertainty. Having sound meta­knowledge means being able to predict within reason­able ranges.

W hether you should focus on 90 percent, 70 per­cent, or just 50 percent confidence ranges depends on

SLOAN MANAGEMENT REVIEW/WINTER 1992

i

r the issues and risks involved. When building a complex We strongly disagree. W hether you know a lot or a little new oil refinery, where the downside risks are high, you about a subject, you are still responsible for knowing may want to incorporate even extreme swings in oil how much you don't know. If you know a lot, as a com-prices. In that case, perhaps a 95 percent confidence puter industry manager should, your 90 percent confi-range on future crude oil price levels should be assessed. dence intervals will be narrow; if you know less, they However, for estimating regional sales levels, you may should be wider. In either case, your subjective 90 per-want a 50 percent range, as you can cope more easily cent confidence intervals should, by definition, capture with surprises outside that range. the true answers 90 percent of the time. (IBM had

We and others have found that whether managers are 365,000 employees on 31 December 1990.) asked for 50 percent, 70 percent, or 90 percent confi- In actuality the job relevance of the questions does dence ranges, few employees are able to supply them re- affect results, possibly because experience reduces over-alistically. Even experts, who by definition know a lot confidence. In Table 1 we see that managers in the com-about a specialized topic, are often unable to express puter firm did better on firm-specific questions (58 per-precisely how much they do not know. Yet to size up cent misses) than on those covering their entire industry and factor uncertainty into our judgments is crucial to (80 percent misses). The data-processing managers, too, successful decision making. Experimental evidence sug- showed less overconfidence on industry-specific items gestS that this is a serious weakness in human judgment, (42 percent misses) than on general business facts (62 even among those well versed in the use of quantitative percent misses). tools. In technical language, few people are well calibrat- Although these results suggest that job relevance tends ed; that is, few people can accurately assess their uncer- to reduce overconfidence, would such a pattern be con-tainty. In business, this translates into risk underesti- firmed by a more systematic study? Moreover, is the re-mates, missed deadlines, and budget overruns. duction in overconfidence only partial, or would ques-

T able 1 summarizes some results we have collect- tions very specific to people's jobs drastically reduce ed from different industries, most often using ques- overconfidence? We tested this using ninety-six profes-tions tailored to that industry and occasionally to a sionals drawn from a variety of occupations. We used specific firm. No group of managers we tested ever two different confidence quizzes: the first contained fif-exhibited adequate metaknowledge; every group be- teen job-specific questions; the second contained fifteen lieved it knew more than it did about its industry or questions unrelated to these professionals' jobs. The un-company. Of the 2,000-plus individuals to whom we related questions were created by having pairs of profes-have given a ten-question quiz using 90 percent con- sionals exchange questionnaires. Thus, one person's job-fidence inter vals, fewer t h a n 1 p e r c e n t w e r e n o t specific questions become the other's unrelated questions overconfident. and vice versa. As a check, we asked everyone afterwards

Our own evidence in Table 1 is confirmed by a large to rate the job relevance of the fifteen questions in the body of similar results from different -�- �–�—-�————�——��—–�—- —– – – —�-

professions, levels of expertise, and Table 1 Overconfidence across Industries ages.2 The only cross-cultural studies, done with Asian managers of several nationalities, further confirm the ubiquity of overconfidence.3

If a question falls outside your area of expertise, should you be excused if your confidence interval misses it? W hen we ask, "How many total em­ployees did IBM have on its payroll on 31 December 1990?" managers outside the computer industry some­times remark that IBM's staff size is

Industry Tested

Advertising

Computers

Data processing

Money management Petroleum

Pharmaceutical Security analysis

Kind of Questions Used in Test

Industry Industry Industry Firm Industry General business Industry Industry & firm Industry & firm Firm Industry

Percentage of Misses Size**

Ideal* Actual

10% 61% 750

50 78 750

5 80 1290

5 58 1290

10 42 252

10 62 261

10 50 480

10 50 850

50 79 850

10 49 390

10 64 497

irrelevant to their job, so they should * The ideal percentage of misses is 1 DO% minus the size of the confidence interval. Thus. a 10% ideal be forgiven for their poor p e r for- means that managers were asked f o r 90% confidence intervals. mance on an overconfidence quiz. -=-��e!�ta���rn

be�o��ucl_��e�n�s-ma� _e a_cro:s�persons �n� ��es���=� ____ _

SLOAN MANAGEMENT REVIEW/WINTER 1992 RUSSO & SCHOEMAKER 9

– I

first quiz, using a scale from 1 (irrelevant) to 7 (highly relevant). For these 90 percent confidence ranges, the unrelated quiz y ielded 53 percent misses (instead of the ideal 10 percent). For the job-relevant quiz, rhe percent­age of misses went from a high of 58 percent for the least relevant questions to 39 percent for the most relevant ones. Figure 1 display s rhis downward trend.' Note that overconfidence does not vanish, but remains at 29 per­cent over rhe ideal, even for rhe most relevant questions.

In sum, better primary knowledge is generally associ­ated with better (though still imperfect) metaknowl­edge. That is, experts know better what they don't know, and this fact is one key to effective solutions, as we discuss next.

Developing Good Metaknowledge

How might professionals develop a sharper sense of how much rhey do and do not know? Once the existence of overconfidence is acknowledged, two elements are es­sential: feedback and accountability.

Feedback rhat is accurate, timely, and precise tells us by how much our estimates missed the mark. Account­ability forces us to confront that feedback, recalibrate our perceptions about primary knowledge, and temper our opinions accordingly.

One mistake we often see managers make is equating experience and learning. Experience is inevitable; learning is not. Overconfidence persists in spite of experience be-

10

Figure 1

Proportion of Misses

Job Relevance of Questions Partially Reduces Overconfidence

.6

.5

.4

.3

.2

. 1

Remaining Overconfidence

Ideal proportion of misses

3 4 5

Relevance

Russo & ScHOEMAKER

6 7 Most

cause we often fail to learn from experience.' In order to learn, we need feedback about the accuracy of our opin­ions and doubts. We also need the motivation to trans­late this information into better metaknowledge.

At least three groups of professionals have used sys­tematic feedback and accountability to develop excellent metaknowledge: Shell's geologists, public accountants, and weather forecasters. • Shell's Geologists. Recall the earlier example of Royal Dutch/Shell, the Anglo-Dutch oil and gas giant. Shell had noticed that newly hired geologists were wrong much more often than their levels of confidence im­plied. For instance, they would estimate a 40 percent chance of finding oil, but when ten such wells were ac­tually drilled, only one or two would produce. This overconfidence cost Shell considerable time and money.

These judgment flaws puzzled senior Shell execu­tives, as the geologists possessed impeccable credentials. How could well-trained individuals be overconfident so much of the time? Put simply, their primary knowledge was much more advanced than their metaknowledge. To develop good metaknowledge requires repeated feed­back, which was coming too slowly and costing too much money.

In response, Shell designed a training program to help geologists develop calibration power. As part of this train­ing, rhe geologists received numerous past cases rhat in­corporated the many factors affecting oil deposits. For each case, they had to provide best guesses as well as ranges that were numerically precise. Then they were given feedback as to what had actually happened. The training worked wonderfully: now, when Shell geologists predict a 40 percent chance of producing oil, four out of ten times rhe company averages a hit. • Public Accountants. When experienced auditors pro­vided estimates and confidence ranges for account bal­ances, they actually proved slightly underconfident.6 Their ranges were too wide rather than too narrow. Per­haps accountants have learned to compensate for over­confidence because of their role as detectors of fraud and error. The profession places an extraordinarily high value on conservative judgments. • Wearher Forecasters. But what, then, are we to make of the only other professional group that has been found to be well calibrated: U.S. Weather Service forecasters? Figure 2 tells a remarkable story of accurate subjective probabilities/ W hen U.S. Weather Service forecasters predicted a 30 percent chance of rain, as they did 15,536 times in this study, it rained almost exactly 30 percent of the time. This superb accuracy holds along the entire range of probability, except at the highest levels. When a

SLOAN MANAGEMENT REVIEW/WINTER 1992

I 00 percent chance of rain is predicted, it actually rains only 90 percent of the time. This prediction error re­flects deliberate caution on the part of the forecasters.

What these three groups have in common is precise, timely feedback about repeated judgments in a field whose knowledge base is relatively stable, unlike the stock market or fashion industry, for instance. Furthermore, all three groups are held accountable by their supervisors or professional colleagues for the accuracy of their confi­dence judgments. Within a day, the weather forecasters receive feedback about whether or not it rains, and their predictive performance is factored into their salary in­creases and promotions. We believe that timely feedback and accountability can gradually reduce the bias toward overconfidence in almost all professions. Being "well cali­brated" is a teachable, learnable skill.

Organizations can accelerate the slow, costly process of learning from experience by keeping better track of managerial judgments and estimates. Performance re­views should emphasize the value to the firm of realism and back this emphasis up with both assessments and incentives. In addition, training programs can provide feedback on simulated or past decisions whose out­comes are not widely known, just as Shell's training pro­gram did.

Systematic feedback works, even though it treats only

Figure 2 U.S. Weather Service Forecasting Accuracy

100

90

80

70

60 Actual

Precipitation 50 Frequency (%) 40

30

20

10

37303 _ _.__….__-L-_.___.____.___…..�_….J�.L….J 0 60 70 80 90 100

Forecaated Probability of Precipitation in the Next 24 Hours (%1

Note: N�mbers beside each point indicate sample sizes. Reprinted by permiSSIOn from A.H. Murphy and R.l. Winkler. "Probability Forecasting 1n Meteorology," Journal of the American Statistical Association 79 (1984): 489-500.

SLOAN MANAGEMENT REVIEW/WINTER 1992

the symptoms of overconfidence. That is, it corrects overconfidence without teaching what caused it in the first place. Several other techniques for reducing over­confidence directly attack its causes. No single cause or prototypical situation can be consistently connected with overconfidence. There are three classes of causes: cognitive, physiological and motivational.

Cognitive Causes of Overconfidence

• Availability. A major reason for overconfidence in pre­dictions is that people have difficulty in imagining all the ways that events can unfold. Psychologists call this the availability bias: what's out of sight is often out of mind.8 Because we fail to envision important pathways in the complex net of future events, we become unduly confident about predictions based on the fewer path­ways we actually do consider.

The limited paths that are evident (e.g., the expected and the ideal scenarios) may exert more weight on like­lihood judgments than they should. Bridge players pro­vide a telling example of how availability can cause over­confidence.9 More experienced bridge players are better­calibrated bidders because they take into account more unusual events or hands. Less experienced players be­lieve they can make hands they often cannot, precisely because they fail to consider uncommon occurrences. • Anchoring. A second reason for overconfidence relates to the anchoring bias, a tendency to anchor on one value or idea and not adjust away from it sufficiently.10 It is typical to provide a best guess before we give a ballpark range or confidence interval. For example, we usually es­timate next quarter's unit sales before we come up with a confidence range. The sales estimate becomes an anchor point and drags the high and low brackets, preventing them from moving far enough from the best estimates.

How strong is the best estimate's pull? To answer this question, we presented twenty trivia questions (e.g., "What is the length of the Nile River?") to two groups of managers. One group (of eighty-four people) first gave a best estimate, that is, an anchor point, and then provid­ed a 90 percent confidence interval. The second group (of fifty-one people) directly supplied a confidence range without ever committing to a best guess. The first group scored 61 percent misses (compared to the ideal of 10 percent). In contrast, the unanchored group's intervals were wider and missed only 48 percent of the true an­swers. Thus, overconfidence was reduced substantially by simply skipping best guesses and moving directly to ranges. (The Nile River is 3,405 miles long.)

Although this de-anchoring technique has yet to be

Russo & SCHOEMAKER 11

verified outside the controlled laboratory, we see no rea­son why it should not work as well in managerial envi­ronments. Interestingly, how well it works will depend on managers' ability to focus on the confidence interval and block out of their thinking any earlier estimate that might serve as an anchor. • The Confirmation Bias. A third cognitive reason for overconfidence concerns our mental search process. When making predictions or forecasts, we often lean to­ward one perspective, and the natural tendency is to seek support for our initial view rather than to look for disconfirming evidence. Unfortunately, the more com­plex and uncertain a decision is, the easier it is to find one-sided support. Realistic confidence requires seeking disconfirming, as well as confirming, evidence.''

How much weight to give to evidence, pro or con, is a complex issue depending both on the strength of the evidence itself and on the credibility of its source. Griffin and T versky, for instance, suggest that people over-weight the strength of evidence (e.g., how well a candidate did in an interview) relative to the credibility of that type of evidence (the limited insight gained from any single interview).12 Whenever source credibility is low, as is often the case in business, and the strength of the evidence is highly suggestive, overconfidence is like­ly to occur. Thus, the interviewed candidate is too con­fidently predicted to be a winner or loser, given the falli­ble, limited evidence obtainable from a short, one-time interview. Ironically, Griffin and T versky predict under­confidence under reverse circumstances, when the credi­bility of the source is high, but the evidence does not point strongly to one action or opinion. • Hindsight. Hindsight makes us believe that the world is more predictable than it really is. What happened often seems more likely afterwards than it did before­hand, since we fail to appreciate the full uncertainty that existed at the time. Recall, for instance, George Bush's landslide victory over Michael Dukakis in 1988 (54 per­cent of the popular vote). At the time of the nominating conventions, the outcome of the election seemed far from certain. Indeed, University of Chicago MBA stu­dents that summer gave Bush only a 49 percent chance of winning,u and the political press frequently cited the "wimp" factor. Nonetheless, the results might seem quite predictable some years later. Hindsight instills an illusion of omniscience.

Cognitive Remedies to Overconfidence

What remedies are available for cognitive sources of overconfidence? We look at five techniques.

12 RUSSO & SCHOEMAKER

• Accelerated Feedback. Recall how successful Shell was in training its junior geologists on past cases where the outcome was known, so they could get immediate feed­back. Simple experiments with trivia questions have demonstrated the efficacy of accelerated feedback.'4 This kind of training can be especially effective for new em­ployees. Using tests derived from actual company records, employees could be trained to estimate their confidence in knowledge relevant to their new jobs. At first, these predictions will almost certainly be overcon­fident, but good feedback will quickly reduce it. And, in contrast to learning from experience, which tends to be slow and expensive, good feedback will reduce overcon­fidence cheaply.

Bur what can you do when faced with a single deci­sion that you must make soon? Try to improve your thinking by bringing to mind relevant considerations that might easily be overlooked. The next four tech­niques offer specific methods for doing so. • Counterargumentation. Think of reasons why your initial beliefs might be wrong, or ask others to offer counterarguments. Several studies have demonstrated the power of generating counterarguments, including one where a major company tested its managers by ask­ing questions such as the following:"

Our company's current liabilities (defined as notes payable, short-term loans, etc.) were $1,911 million and $1,551 million as of December 31, 19xx, and March 31, 19xx, respectively. For October 31, 19xx, the company's current liabilities will be (circle one):

(a) greater than $1,900 million

(b) less than or equal to $1,900 million

Give your subjective probability that you will be correct:__ %.

Half of the participants in the experiment were merely asked to circle (a) or (b) and then state how confident they were about their choice. This group's mean estimat­ed probability of being correct was 72 percent. However, they actually picked the correct answer only 54 percent of the time. Hence, they were overconfident by 18 percent. The other half of the participants were asked to think of "the major reason why the alternative circled might be wrong" before giving their subjective probabilities. That is, they were asked to think disconfirmingly and provide at least one counterargument to their initial guess. Then they were allowed to change their answer if they wished. This group's average estimated probability of being

SLOAN MANAGEMENT REVIEW/WINTER 1992

correct was 73 percent, and they actually picked the cor­rect answer 62 percent of the time. Thus, their level of overconfidence was only 11 percent, a reduction of nearly two-fifths thanks to a single counterargument. Other studies have found that, when listing pros and cons, the cons do the most good in countering overconfidence.'(,

But is this practical? It depends. We see no reason why major capital budgeting requests could not have a counterargument section in which managers are asked to identify the major reasons not to go ahead. And if the project does fail later, the actual causes had better be listed in this contrarian section of the report. A warning, however. To be useful, this process must be taken seri­ously, with serious consequences (both good and bad) for the managers involved. Otherwise it may degenerate into a useless formality or, worse, corporate "gaming" (i.e., managers withholding their genuine concerns in favor of saying whatever gets the budget approved). A second problem is that managers may truly not recog­nize potential hazards. The next tactic, Paths to Trouble, addresses that problem. • Paths to Trouble. If we are overconfident in predict­ing success because we cannot see the paths to potential trouble, fault trees may help. A fault tree is a hierarchi­cal diagram designed to help identify all the paths to some specific "fault" or problem. For an example, see Figure 3. To be useful, fault trees must be reasonably complete, at least in identifying the major categories of potential trouble. If they are not, chances are that even specialists will fail to realize what is missing. 17 People as­sume the causes listed account for almost every thing that could go wrong and underestimate the final catego­ry, "all other" causes of failure. In the study that used the restaurant fault tree, the "all other" category was es­timated by hospitality industry managers to contain only 7 percent of the chance that something might go wrong, whereas it really contained 54 percent of the chance. 1M

How can this blindness be overcome? Warning peo­ple does not seem to help.19 What does work is asking people to extend the fault tree by listing additional caus­es of the problem. 20 In the restaurant study, some people were shown branches with only six instead of twelve causes and asked to provide further reasons. When two more causes were listed by managers themselves, the original omission error of 46 percent dropped to 23 percent; when four were listed, it dropped to 12 per­cent; and when six possible causes were added, it disap­peared entirely. In sum, the more causes generated, the smaller is the error of assuming that all relevant causes are already listed.

SLOAN MANAGEMENT REVIEW/WINTER 1992

• Paths to the Future. If deeper thinking is called for, bey ond the listing of reasons, explicit scenario analysis may be useful. Whereas fault trees highlight individual causes, scenarios focus on their conjunction. Scenarios are script-like narratives that paint in vivid detail how the future might unfold in one or another direction. Envisioning vastly different worlds than those expected has helped companies like Royal Dutch/Shell to better estimate economic and political uncenainty.2' A direct test that compared 90 percent confidence intervals be­fore and after scenario construction found, on average, a 30 percent stretching of ranges. 22 Asking managers to construct different scenarios makes them better appreci­ate the uncertainty in key parameters or estimates. In addition, it often provides new ideas for innovation or competitive positioning. • Awareness Alone. Although these techniques are valu­able, we happily acknowledge that, for many managers, awareness alone may be all that is needed. Good man­agers often devise their own solutions to the problems of overconfidence.

Recall the head loan officer who was certain that he and his staff knew their competitors quite well. Mter failing the tailored overconfidence quiz, he took im­mediate action. Each officer was required to con­tribute information to a "competitor alert" file. And each was required to check the file weekly to gain a m o r e realistic a p p r e c i a t ion of their competition. Within three weeks, a loan officer found information in the file signaling that a major client was contem­plating a shift to another bank. The competitor was not one of the city's major commercial banks, and the magnitude of the business at risk exceeded the com­petitor's legal lending limit. However, by joining with another institution, it was able to offer a loan large enough to meet the client's needs. Thus alerted, the loan officer in charge of the account convinced the client not to switch banks, saving $160,000 in annual revenue.

The head of sales for Index Technology took a differ­ent approach when his salespeople were overconfident about if and when orders of the company's product would be written by potential customers. He called some customers himself. His salespeople didn't like it, but the approach worked: soon they were predicting orders and the timing of those orders much more accurately.

A negotiation experiment further underscores the value of awareness alone. 2·' Subjects believed they had a 65 percent chance of winning in a simulated negotia­tion task entailing binding arbitration. In this set-up, for every winner there had to be a loser, implying a 50

Russo & ScHOEMAKER 13

Figure 3 Fault Tree for Restaurant Failure

I Restaurant failure due to decreasing profits j I

I I

Decreasing revenues I Increasing costs I I I

I I I I I

Decreasing number Decreasing average Increasing food costs Increasing labor costs Increasing overhead

of customers food/beverage check cost

I I I I I •Incorrect pricing • Decreasing perceived • Improper purchasing and • Increasing overtime • High debt service cost

• Unclear image of value by customers receiving scheduling • Poor facility design

property • Incorrect pricing • Menu variety !too limited • Decreasing productivity • High occupancy costs

• Changing atmosphere • Inadequate service pace or too extensive) • Poor organizational •Inadequate capital

• Outdated restaurant • Poor merchandising • Poor sales forecasting climate structure

concept • Lack of employee • Changes in supplier • Union rules • Low sales volume

• Inadequate promotion motivation market • Improper physical layout •Improper growth rate

and advertising • Changing customer • High waste and leftovers and equipment • High administrative costs

• Changing customer dining out budget • Inadequate number of • Poor employee selection • Incorrect work method of

expectations • Changing mix of food/ employees and training personnel

• Poor food quality beverage sales • High theft • High employee turnover •Improper business hours

• Changing competition •Inconsistent food quality • Poor supervision • Menu variety too • Outdated facility and

• Poor service quality • Changing competition •Incorrect size of food extensive equipment • Changing customer • Improper portion size portions • Poor supervision • High inflation

demographics • Changing customer • Improper storage and • Inadequate wage • Poor credit rating • Lack of menu variety demographics issuing structure •All other • Changing location • Changing customer •All other •Inefficient employee

characteristics tastes scheduling • All other • All other • Increasing benefit costs

• All other

Reprinted by permission of John Wiley & Sons. Ltd .• copyright 1988. From l. Dube-Rioux and J.E. Russo, "An Availability Bias in Professional Judgement," Journal of Behavioral Decision Making 1 11988): 223-237.

percent chance of winning. Hence, most of the people were overconfident. The researchers then took a ran­dom half of the people aside and warned them about overconfidence. Compared to the unwarned group, those forewarned were 30 percent more likely to reach a negotiated agreement instead of having to turn to costly arbitration, and they achieved net dollar benefits that were 70 percent higher. As skilled lawyers know, negotiating a settlement is one area where realism pays.

General Versus Specific Awareness

Although general awareness of a bias is invaluable, it does not guarantee that the bias will be spotted in every in­stance. Consider this study. Twelve financial officers were asked to estimate ten quantities pertinent to their

14 Russo & SCHOEMAKER

organization's business operations and to provide a 90 percent confidence interval for each.24 As usual, these in­tervals failed to capture the true value a high percentage of the time; in this case, the failure rate was 78 percent versus the ideal of 10 percent. In addition, the financial officers were asked to estimate " how many of the ten in­tervals you gave … will contain the actual value." This was asked immediately after the ten intervals were con­structed and before the true answers were revealed. Of course, every officer should have answered "nine" be­cause that is what a 90 percent confidence interval means, by definition. However, only one did. The oth­ers estimated that fewer than nine of the ten intervals would capture the true value. On average, the twelve of­ficers guessed that they would miss 5.6 of the ten ques­tions but couldn't tell which ones. These data suggest that people are more aware of overconfidence in general

SLOAN MANAGEMENT REVIEW/WINTER 1992

than they are in particular. The same problem – of general awareness but specif­

ic blindness – was described by John Stuart Mill, the 19th century economist and social philosopher, in On Liberty:

Unfortunately for … mankind, the fact of their fal­libility is for from carrying the weight of their practi­cal judgment, which is always allowed to it in theo­ry; for while everyone knows himself to be fallible, few . . . admit the supposition that any opinion of which they feel very certain may be one of the exam­ples of the error to which they acknowledge them­selves to be liable.

Physiological Causes of Overconfidence

Because overconfidence is a distortion of judgment, it is often thought of as a purely mental phenomenon; how­ever, at times it has biochemical causes. Euphoria, the elated feeling of well-being that commonly follows per­sonal or professional success, may cause overconfidence. (The biochemical compounds involved in euphoria ap­pear to be hormones, such as adrenalin and endor­phines, that the body produces as a response to strong emotional reactions.) We also suspect that drugs like co­caine and alcohol can produce overconfidence. 25

Ford Motor Company provides an example of how a major firm dealt with the negative side effects of eu­phoria. As the 1970s ended, Ford faced hard times: re­duced market share, layoffs, and the superior quality of Japanese cars. In response to these conditions, Ford organized meetings of its plant managers and assistant managers to solicit and communicate suggestions for improving manufacturing quality. The flood of ideas, and the resulting enthusiasm for what might be ac­complished, s w ept n e a r l y e v e r yone a way. W i s e l y, Ford's top management imposed a cooling-off period of several weeks before the returning managers could implement any of the suggestions. Because of the eu­phoric mood at the end of the meetings, senior execu­tives distrusted their managers' judgment, and they wanted time for a more calculated look prior to com­mitting major funds.

Dealing successfully with physiological causes of overconfidence, as with all types of overconfidence, re­quires awareness of the problem: you can't fix it if you can't find it. In this regard, individual awareness is the single most important factor. If you are euphoric, wait

SLOAN MANAGEMENT REVIEW/WINTER 1992

to commit yourself to a plan of action, just as, if you drink, don't drive.

Overconfidence in Group Judgments

By this time, you may wonder if groups do better than individuals when sizing up uncertainty. The answer is mixed.26 Group judgments can be better than individual ones, precisely because in groups people are forced to recognize that others see the world differently than they do. This often sparks a realization that perhaps their own views are held with unjustifiable conviction. A t other times, however, groups may bolster the majority opinion to even more extreme levels.

To test group overconfidence, we conducted a simple experiment with eighty-three managers. First, people were asked to privately form 90 percent confidence ranges on ten questions. Then they were asked to com­pare and discuss their ranges in groups of three or four in order to come up with a single group range for each question. We did not specifY how this was to be accom­plished. Some groups argued heatedly; others merely av­eraged the individual guesses; still others used the most extreme values in the group as their outer brackets. After the group decisions were made, people were al­lowed to change their private ranges. The initial, unre­vised private judgments generated an average of 72 per­cent misses (compared to an ideal of 10 percent), signifying serious overconfidence. Group judgments were significantly less overconfident, with 56 percent misses on average. The revised private judgments made after the group decisions resulted in 62 percent misses.

Making Group Judgments Better

On average, the group judgments were better than indi­vidual judgments in the above task. At worst, they forced a compromise; at best, they encouraged openmindedness. Individually, however, people may still anchor too strong­ly to their initial view and return to it when given the chance. This stubbornness can be to their, and their com­pany's, detriment. There are relatively simple techniques for minimizing this recidivist tendency.

Delphi techniques and other procedures for sharing and averaging opinions are especially feasible in a net­worked PC environment. Rather than go around the table, collect people's initial estimates and ranges pri­vately. Next, share these ranges and only then com­mence a debate. Mter group discussion, ask for one

RUSSO & SCHOEMAKER 15

more round of opinions and run with those averages. An extensive literature exists on expert aggregation and averaging. as well as experimental work on individual versus group calibration.!-

Motivational Factors in Overconfidence

Overconfidence isn't all bad! One legitimate cause of overconfidence is our need to believe in our abilities. Indeed, confidence in one's abilities is particularly wide­spread. At the beginning of a course, we often ask our MBA students (anonymously) whether their final grade will be in the bottom or top half of the class. The great majority are certain they will finish in the top half, and they are willing to bet on it.

Many of these people are distorting reality, yet their optimism has motivational value. Would risky projects be undertaken if a few key people did not have an unre­alistic belief in their chances of success? As Goethe wrote, "For a man to achieve all that is demanded of him he must regard himself as greater than he is."

If the motivating value of overconfidence is clear, so is its downside. The value and danger of overconfidence may especially conflict for entrepreneurs. They often take risks others would not, and they must persuade in­vestors and employees to join them in highly uncertain endeavors. Yet their eventual success also requires real­ism. A partner at a venture capital firm summarized this problem: "You expect entrepreneurs to have … an un­shakable sense that they absolutely cannot fail. Yet since we will be partners with these people, we want to be sure that their egos will not stand in the way of making the best decisions for the business."

Moreover, to succeed in many business endeavors, we have to project confidence even when it cannot be justi­fied. Because people often equate confidence with com­petence, you had better sound confident if you want your opinions to be treated as credible. It is seldom easy to stand up at an important meeting and say, ''I'm not sure." Instead, people go out on a limb.

Can anything be done to reconcile the danger of dis­torting reality with the value of optimism? Perhaps the best advice is: Don't fool yourself. Don't permit yourself to be overconfident when making important decisions or commitments.

Deciding and Doing

We believe that much of the damage can be avoided if managers distinguish between deciding and doing. Deciding requires realism. But in implementing the

16 Russo & SCHOEMAKER

decision, the motivational benefits of overconfidence frequently outweigh its dangers.

Separating deciding from doing is not simple, and it's harder than it used to be. A century ago, business orga­nizations were more like vertical pyramids: the deciders were on top and the workers were underneath. But in today's "flattened" organizations, every manager is both a decider and a doer.

So what should today's managers do? Our recom­mendation is to be aware of when you are fUnctioning as a decider and when your primary role is that of doer, motiva­tor, or implementer. When you are deciding, be realistic, both about how much you know and how much you don't know. W hen you are implementing, indulge in overconfidence when, and if, it is valuable to your per­formance or that of others.

All of us need self-confidence to function. We might not show up for work every morning if we did not believe we could make a difference. Nevertheless, too much con­fidence can backfire – can cause us to bet on plans, peo­ple, or projects which a more realistic appraisal would have rejected. Though normally an advocate of rational calculation, Lord Keynes keenly observed this human dilemma: "A large proportion of our positive activities de­pend on spontaneous optimism rather than on mathe­matical expectation … if animal spirits are dimmed and the spontaneous optimism falters, leaving us to depend on nothing but mathematical expectation, enterprise will fade and die."+

References

The authors acknowledge janet Sniezek and Ilan Yaniv for their construc­tive comments and jack B. Williams for his editorial advice. I. Linguists distinguish between language competence (the ability to produce coherent statements) and metalanguage (the ability t o state the rules of the language). Such a clear distinction does not always exist between primary knowledge and metaknowledge. Early in the century, U.S. Weather Service forecasters simply predicted whether or not it would rain (a statement of their primary knowledge). Now they provide an explicit probability of rain, making uncertainty assessment an explicit part of their primary knowledge. 2. S. Lichtenstein, B. Fischhoff, and L.D. Phillips, ''Calibration of Probabilities: The State of the Art to 1980," in judgment under Uncertainty: Heuristics and Biases, eds. D. Kahneman, P. Slovic, and A. Tversky (New York: Cambridge University Press, 1982), pp. 306-334. 3. G.N. Wr i g h t a n d L.D. P h i l l i p s , "C u l t u r a l Va r i a t io n s i n P r o b a b i l i s t i c Thinking: A l t e r n a t i v e W a y s o f D e a li n g w i t h Uncertainty," lnternationaljoumal o f Psychology 1 5 (1980): 239-257. 4. All claims made about differences or trends are statistically signifi­cant at the .05 level or lower. The sample sizes for the percentages in Figure 1 range from a low of 122 when relevance = I to a high of 270 when relevance= 7, with the unrelated percentage based on all 1,440 unrelated questions.

SLOAN MANAGEMENT REVIEW/WINTER 1992

5. J.E. Russo and P.J.H. Schoemaker, Decision Traps (New York: Simon and Schuster, 1990). 6. L.A. Tomassini et al., "Calibration of Auditors' Probabilistic Judgments: Some Empirical Evidence," Organizational Behavior and Human Pnfonnance 30 ( 1982): 391-406. 7. A.H. Murphy and R.L. Winkler, "Probability Forecasting in Meteorology," journal of the American Statistical Association 79 ( 1984) :489-500. 8. A. Tversky and D. Kahneman, "Availability: A Heuristic for Judging Frequency and Probability," Cognitive Psychology 4 (1973): 207-232; B. Fischhoff, P. Slovic, and S. Lichtenstein, "Fault Trees: Sensitivity of Estimated Failure Probabilities to Problem Representation," journal of Experimental Psychology: Human Perception and Pnformance 4 ( 1978): 330-344. 9. G. Keren, "Facing Uncertainty in t h e Game of Bridge: A Calibration Study," Organizational Behavior and Human Decision Processes 39 ( 1987): 98-114. 10. P. Slovic and S. Lichtenstein, "Comparison of Bayesian and Regression Approaches to the Study of Information Processing in Judgment," Organizational Behavior and Human Performance 6 (1971): 641-744; A. Tversky and D. Kahneman, "Judgment under Uncertainty: Heuristics and Biases," Science 185 (1974): 1124-1131. II. J. Klayman and Y.W. Ha, "Confirmation, Disconfirmation, and Information in Hypothesis Testing," Psychological Review 94, 2 (1987): 211-228. 12. D. Griffin and A. Tversky, "The Weighing of Evidence and the Determinants of Confidence" (Waterloo, Ontario: University of Waterloo, working paper, 1991). 13. P.J.H. Schoemaker, "Scenario Thinking" (Chicago: Graduate School of Business, University of Chicago, working paper, 1991). 14. For a review, see Lichtenstein, Fischhoff, and Phillips (1982). 15. J. Mahajan and ).C. Whitney, Jr., "Confidence Assessment and the Calibration of Probabilistic Judgments in Strategic Decision Making" (Tucson: University of Ariwna, working paper series #12, 1987). 16. S.J. Hoch, "Availability and Inference in Predictive Judgment," Journal of Experimental Psychology: Learning. Memory, and Cognition I 0 (1984): 649-662. 17. Fischhoff, Slovic, and Lichtenstein (1978). 18. L. Dube-Rioux a n d J.E. R u sso, "An Availability Bias i n Professional Judgment," journal of Behavioral Decision Making I (1988): 223-237. In this study, six of the twelve listed causes in a branch of a fault tree (see Figure 3) were removed. If people, in this case hospitality industry managers, were properly aware of all the major causes, then all of the probability of these six unlisted causes should show up in the last, "all other" category. In fact, very little did, strongly suggesting that what is out of sight is out of mind; i.e., the availability bias operates. 19. Fischhoff, Slovic, and Lichtenstein (1978). 20. Dube-Rioux and Russo (1988). 21. P. Wack, "Scenarios: Uncharted Waters Ahead," Harvard Business Review, September-October 1985, pp. 73-89; P. Wack, "Scenarios: Shooting the Rapids," Harvard Business Review, November-December 1985, pp. 139-150. 22. Schoemaker (1991). 23. MA. Neale and M.H. Bazerman, "The Effects of Framing and Negotiator Overconfidence on Bargaining Behavior and Outcomes," Acadnny ofManagemmt journal28 (1985): 34-49.

SLOAN MANAGEMENT REVIEW/WINTER 1992

24. ).A. Sniezek and T. Buckley, "Level of Confidence Depends on Level of Aggregation," journal of Behavioral Decision Making4 (1991): 263-272. 25. We wonder how many traffic fatalities are caused by alcohol-in­duced overconfidence. Ccnainly driving skills are impaired by alcohol. but this may be only part of the story. A more deadly aspect is that the drinker's confidence is not reduced nearly as much as the ability itself. This confidence gap between the skill levels drivers believe they possess and the reduced levels they actually have seems to be a primary prob­lem with drunk drivers. 26. Despite a presumption that "two heads are better than one," groups do not always make better decisions than individuals. The phe­nomenon known as groupthink is one serious problem. Whether groups are superior seems to depend on whether conflict is articulated or swept under the rug. See: I.L. Janis, Groupthink, 2nd ed. (Boston: Houghton Mifflin, 1982.) 27. R.T. Clemen and R.L. Winkler, "Unanimity and Compromise among Probability Forecasters," Management Science 36 (1990): 767-779; and )A. Sniezek and RA. Henry, "Accuracy and Confidence in Group ) udgment," Organizational Behavior and Human Decision Processes 43 (1991): 1-28.

The answers to the Confidence Quiz: (I) 96,727 patents; (2) 111 Japanese corporations; (3) 59,130,007 arrivals and depanures; (4) 2,076,713; (5) 67,496 degrees; (6) 6,700 deaths; (7) 5,757 miles; (8) 77.8 million units; (9) 147,324 automobiles; and (10) $354 billion.

Reprinted by Permission Sloan Management Review e 1992.

Reprint 3321

Russo & SCHOEMAKER 17