Distilling the Wisdom of Crowds: Prediction Markets vs. Prediction Polls

Abstract

We report the results of the first large-scale, long-term, experimental test between two crowdsourcing methods: prediction markets and prediction polls. More than 2,400 participants made forecasts on 261 events over two seasons of a geopolitical prediction tournament. Forecasters were randomly assigned to either prediction markets (continuous double auction markets) in which they were ranked based on earnings, or prediction polls in which they submitted probability judgments, independently or in teams, and were ranked based on Brier scores. In both seasons of the tournament, prices from the prediction market were more accurate than the simple mean of forecasts from prediction polls. However, team prediction polls outperformed prediction markets when forecasts were statistically aggregated using temporal decay, differential weighting based on past performance, and recalibration. The biggest advantage of prediction polls was at the beginning of long-duration questions. Results suggest that prediction polls with proper scoring feedback, collaboration features, and statistical aggregation are an attractive alternative to prediction markets for distilling the wisdom of crowds.
This paper was accepted by Uri Gneezy, behavioral economics.

Full text

MANAGEMENT SCIENCE
Articles in Advance, pp. 1–16
ISSN 0025-1909 (print) ISSN 1526-5501 (online) http://dx.doi.org/10.1287/mnsc.2015.2374
© 2016 INFORMS
Distilling the Wisdom of Crowds:
Prediction Markets vs. Prediction Polls
Pavel Atanasov
University of Pennsylvania, Philadelphia, Pennsylvania 19104; and Polly Portfolio, Inc., New York, New York 10007,
pdatanasov@gmail.com
Phillip Rescober
Good Judgment, Inc., Philadelphia, Pennsylvania 19103, rescober@goodjudgment.com
Eric Stone
Pluralsight, Inc., Boston, Massachusetts 02118, theericstone@gmail.com
Samuel A. Swift
Betterment, Inc., New York, New York 10010, samswift@gmail.com
Emile Servan-Schreiber
Lumenogic, Inc., 75013 Paris, France, emile@lumenogic.com
Philip Tetlock
Department of Psychology, University of Pennsylvania, Philadelphia, Pennsylvania 19104, tetlock@wharton.upenn.edu
Lyle Ungar
Department of Computer and Information Science, University of Pennsylvania, Philadelphia, Pennsylvania 19104,
ungar@cis.upenn.edu
Barbara Mellers
Department of Psychology, University of Pennsylvania, Philadelphia, Pennsylvania 19104, mellers@wharton.upenn.edu
W
e report the results of the first large-scale, long-term, experimental test between two crowdsourcing methods:
prediction markets and prediction polls. More than 2,400 participants made forecasts on 261 events over
two seasons of a geopolitical prediction tournament. Forecasters were randomly assigned to either prediction
markets (continuous double auction markets) in which they were ranked based on earnings or prediction polls
in which they submitted probability judgments, independently or in teams, and were ranked based on Brier
scores. In both seasons of the tournament, prices from the prediction market were more accurate than the simple
mean of forecasts from prediction polls. However, team prediction polls outperformed prediction markets when
forecasts were statistically aggregated using temporal decay, differential weighting based on past performance, and
recalibration. The biggest advantage of prediction polls was at the beginning of long-duration questions. Results
suggest that prediction polls with proper scoring feedback, collaboration features, and statistical aggregation are an
attractive alternative to prediction markets for distilling the wisdom of crowds.
Keywords: forecasting; prediction markets; crowdsourcing, belief elicitation
History : Received September 14, 2014; accepted September 14, 2015, by Uri Gneezy, behavioral economics.
Published online in Articles in Advance.
1. Introduction
We examine two crowd-based approaches for collecting
predictions about future events: prediction markets
and prediction polls. In an idealized prediction market,
traders are motivated by market profits to buy and sell
shares of contracts about future events. If and when
they obtain relevant information, they act quickly in
the market. Knowledge is continuously updated and
aggregated, making prices generally good estimates of
the chances of future events. In the strong form of the
efficient markets hypothesis, no additional information
should be able to improve the accuracy of the last
price. This form of crowdsourcing is used in many
organizations, bridging the gap between economic
theory and business practices (Cowgill and Zitzewitz
2015, Spann and Skiera 2003, Surowiecki 2005).
Direct probability elicitation methods are quite dif-
ferent from prediction markets and are often misunder-
stood. We examine a version of probability elicitation
that we call prediction polls. Participants offer proba-
bilistic forecasts, either independently or as members
of a team, and update their beliefs throughout the
1

Atanasov et al.: Prediction Markets vs. Prediction Polls
2Management Science, Articles in Advance, pp. 1–16, © 2016 INFORMS
duration of each question as often as they wish. They
receive feedback about their performance using a
proper scoring rule and have many opportunities
for learning. Prediction polls should not be confused
with opinion polls. In opinion polls, respondents are
typically asked about their personal preferences or
intentions, on a single occasion. In prediction polls,
respondents place predictions on future events, aiming
for accuracy. Past research shows that that participants’
predictions of election outcomes provide more accurate
estimates than participants’ stated preferences about
election outcomes (Rothschild and Wolfers 2010).
What is the value of better predictions? The answer
varies across businesses, governments, and individuals,
but revenues in the forecasting industry are estimated
at $300 billion in current dollars (Sherden 1998). Better
predictions are financially consequential. Thus, it is
important to know which methods provide more accu-
rate estimates. Prediction markets are often used by
organizations to answer questions about geopolitical
events, project completion, or product sales. But, in
this context, prediction polls are less common. There is
little experimental evidence on the relative accuracy
of alternative methods. The current paper addresses
this knowledge gap and provides lessons about how
to improve the accuracy of crowdsourced predictions
more generally.
There are theoretical and empirical reasons to believe
that prediction markets and prediction polls are both
valid means of gathering and aggregating information
(Gürkaynak and Wolfers 2006, Hanson 2003). These
methods share essential properties: they account for
information revelation over time, place higher weights
on better forecasters, and use past forecasts to improve
calibration. We put prediction markets and prediction
polls in a head-to-head competition and measured the
accuracy of forecasts they produced over two years of
a geopolitical forecasting tournament.
2. Background and Hypotheses
An opportunity to test the accuracy of prediction
markets against prediction polls arose in 2011 when
the Intelligence Advanced Research Project Agency
(IARPA), the research wing of the intelligence commu-
nity, sponsored a multiyear forecasting tournament.
Five university-based programs competed to develop
the most innovative and accurate methods possible to
predict a wide range of geopolitical events. Will any
country officially announce its intention to withdraw
from the Eurozone before April 1, 2013? Will the six-
party talks on the Korean Peninsula resume before
January 1, 2014? What will be the lowest end-of-day
price of Brent Crude Oil between October 16, 2013, and
February 1, 2014? Our group experimentally tested a
variety of elicitation and aggregation methods. Our
methods produced highly accurate forecasts and helped
our team, the Good Judgment Project, to win the fore-
casting tournament (see the online appendix, available
as supplemental material at http://dx.doi.org/10.1287/
mnsc.2015.2374).
The current paper offers the most thorough analysis
of prediction market data from the Good Judgment
Project, as well as a detailed description of aggregation
challenges and solutions in prediction polls. Our main
focus is the comparison of prediction markets and
prediction polls.
We build on our understanding of individual fore-
casting behavior described in prior work. Mellers
et al. (2014) provide an overview of the forecast-
ing tournament and discuss the positive impacts of
three behavioral interventions—training, teaming, and
tracking—on individual performance in prediction
polls. Mellers et al. (2015a) explore the profiles of
individuals using dispositional, situational, and behav-
ioral variables. Mellers et al. (2015b) document the
performance of the most accurate performers, known
as superforecasters, and present four complementary
reasons for their success. Aggregation techniques for
prediction poll data are discussed in three papers.
Satopää et al. (2014a) offer a simple method for com-
bining probability estimates in log-odds space. The
method uses differential weighting, discounting, and
recalibration or “extremizing” of forecasts to reflect
the amount of overlapping information of individual
opinions. Satopää et al. (2014b) describe a time-series
model for combining expert estimates that are updated
infrequently. Baron et al. (2014) provide a theoretical
justification and empirical evidence in favor of trans-
forming aggregated probability predictions toward the
extremes. Atanasov et al. (2013) develop a method for
aggregating probability estimates in prediction markets
when probabilities are inferred from individual mar-
ket orders and combined using statistical aggregation
approaches. Finally, Tetlock et al. (2014) discuss the role
that tournaments can take in society by both increasing
transparency and improving the quality of scientific
and political debates by opening closed minds and
increasing assertion-to-evidence ratio.
We begin by describing prediction markets and
prediction polls. We then discuss the aggregation chal-
lenges inherent in dynamically elicited probability
forecasts. Next, we offer hypotheses and test them in
two studies covering different seasons of the tourna-
ment. We also examine the sensitivity of our results to
choice of aggregation parameters, scoring rules, and
potential moderators.
2.1. Prediction Markets
Since the work of Hayek (1945), economists have
championed the power of markets as a mechanism for
aggregating information that is widely dispersed among

Atanasov et al.: Prediction Markets vs. Prediction Polls
Management Science, Articles in Advance, pp. 1–16, © 2016 INFORMS 3
economic actors. The efficient markets hypothesis posits
that the latest price reflects all information available to
market participants, so future price movements are
unpredictable without insider knowledge (Fama 1970).
Prediction markets build on this rich intellectual
tradition and are designed to produce continuously
updated forecasts about uncertain events. In binary
options markets, traders place bets on the probability
of event occurrence. For example, an election contract
may pay $1 if candidate A is elected president, and $0
otherwise. The latest price is presumed to reflect the
current best guess about the probability of the event.
A contract trading at 60 cents on the dollar implies that
market participants, in aggregate, believe the event
is 60% likely to occur. Those who think the current
price is a poor estimate of the event probability have
incentives to trade.
Evidence suggests that prediction markets can out-
perform internal sales projections (Plott and Chen 2002),
journalists’ forecasts of Oscar winners (Pennock et al.
2001), and expert economic forecasters (Gürkaynak
and Wolfers 2006). Furthermore, prediction markets
perform at least as well as aggregated opinion polls of
voter preferences (Berg et al. 2001, Rothschild 2009).
Some researchers have noted systematic distortions
in market prices, including the favorite long-shot bias
(Page and Clemen 2012) and partition dependence
(Sonnemann et al. 2013). Theoretical work has described
conditions under which prediction markets fail to
incorporate individual beliefs (Manski 2006, Ostrovsky
2012). However, these objections have not changed the
overall impression that prediction markets are superior
to most other prediction methods. Such results have
led Wolfers (2009) to recommend that “rather than
averaging forecasts of
000
[economic forecasters, they]
should be put in a room and told to bet or trade against
each other” (p. 38).
In the current study, several hundred participants
made trades in a continuous double auction (CDA)
market. In a CDA, buyers and sellers are matched to
determine the market price. Traders place orders in
the form of bids (buy orders) and asks (sell orders),
and records are kept in an order book. Trades are
executed when an order is placed and the highest
buying price matches or exceeds the lowest selling
price. Rules of the tournament prohibited us from
using real monetary incentives so all transactions were
made using hypothetical currency.
2.2. Prediction Polls
In prediction polls, forecasters express their beliefs by
answering the question, “How likely is this event?”
Probability judgments are validated against ground
truth using a proper scoring rule. Forecasters receive
feedback on their performance, in much the same
way that prediction market traders are ranked based
on earnings. A similar method has appeared in the
literature under the name of competitive forecasting
(Servan-Schreiber 2007).
There are two important distinctions between predic-
tion polls and other polls or surveys. First, in prediction
polls, participants are asked for probabilistic forecasts,
rather than preferences or voting intentions. Forecasts
are elicited in a dynamic context. Forecasters update
their predictions whenever they wish, and feedback is
provided when events are resolved. Second, forecasters
compete against other forecasters. Servan-Schreiber
(2012) has argued that competitive features encourage
better search processes and more accurate inferences.
Third, prediction polls rely on crowds with dozens,
hundreds, or thousands of individuals who may be
knowledgeable but are not necessarily subject mat-
ter experts, which distinguishes polls from expert
elicitation techniques.
Our team examined two forms of prediction polls.
In the independent poll condition, forecasters worked
independently and had no access to the opinions of
others. In the team poll condition, forecasters worked
online in teams of approximately 15 members (see
Mellers et al. 2014). They shared information, discussed
rationales, and encouraged each other to forecast. Teams
did not need to reach consensus; they made individual
predictions, and the median was defined as the team
forecast. Individual leader boards (within teams and
across independent forecasters) displayed Brier scores.
Team leader boards (across teams) displayed group
performance using Brier scores. Forecasters in team
prediction polls reliably outperformed forecasters in
independent prediction polls.
2.3. Comparisons of Markets and Polls
Studies comparing prediction markets and methods
similar to prediction polls have used observational field
data or small-scale laboratory experiments. Most have
focused on short-term sports predictions or general
knowledge tests. In contrast, we examine geopolitical
predictions on questions that are open for several
months, on average.
Chen et al. (2005) compared the accuracy of linear
and logarithmic aggregations of polls
1
to prices in a
test based on final scores of football games. Accuracy of
the methods did not differ using the absolute distance
rule, the quadratic scoring rule, or the logarithmic
scoring rule.
2
Goel et al. (2010) compared real-money
1
In the linear polls, forecasts were aggregated with an arithmetic
mean. In the logarithmic polls, forecasts were aggregated with a
geometric mean.
2
Using a similar data set, Servan-Schreiber et al. (2004) found
prediction markets to be more accurate than survey-based forecasts
(from ProbabilityFootball), but the difference in results was traced to
the imputation of missing forecasts. Servan-Schreiber et al. (2004)
imputed 50% probability values for missing forecasts, which resulted
in higher errors. Chen et al. (2005) omitted missing observations.

Atanasov et al.: Prediction Markets vs. Prediction Polls
4Management Science, Articles in Advance, pp. 1–16, © 2016 INFORMS
prediction markets to probabilities from opinion polls
(ProbabilitySports and Amazon’s Mechanical Turk).
Prediction markets were more accurate, but not by a
significant margin.
Rieg and Schoder (2010) found no differences in
accuracy when comparing markets and opinion polls
in small-scale experiments, where each of 6 to 17 par-
ticipants made a single prediction in either a market or
a poll. In another small-group experiment, Graefe and
Armstrong (2011) compared the accuracy of answers to
general knowledge questions using the Delphi method,
the Nominal Groups technique, and a prediction market
with a logarithmic market scoring rule (LMSR) (Hanson
2003). The accuracy of the methods was approximately
equal. In summary, prior research has found polls and
markets to be closely matched in accuracy. None of the
existing studies simultaneously featured experimental
assignment and evaluation on a large set of long-term,
real-world questions.
2.4. Forecast Aggregation in Prediction Polls
An important challenge for dynamic forecasting is
aggregating forecasts over time (Armstrong 2001).
We discuss a parsimonious and effective aggregation
algorithm, which our team devised to address three
challenges—outdated predictions, heterogeneity of
individual skills, and miscalibration—discussed below.
(1) Outdated Predictions. When forecasts are continu-
ously elicited and updated over time, more recent ones
usually contain fresher, more valuable information.
For example, one-day-ahead temperature forecasts are
highly accurate, whereas those issued 10 days prior are
as accurate as predictions based purely on historical
averages (Silver 2012).
However, although the single most recent forecast is
likely to be closer to the truth than an older one, just
one forecast can be noisy and inaccurate. The challenge
is to strike the right balance between recency and
number of forecasts. In prediction markets, the last price
is treated as the crowd’s best estimate. In CDA markets,
two marginal traders produce the price: the most
recent highest bidder and lowest-price seller, although,
in highly liquid markets these two individuals have
limited ability to influence the price. If traders had
chosen to transact at different prices, others could have
jumped in to set the price. The availability of price
history and the depth of the order book make it easier
to maintain stable prices.
In prediction polls, forecasters have incentives to
update their predictions, but they do not all update at
the same time. A simple mean relies too much on older
predictions, making it unresponsive to new develop-
ments. One solution is exponential discounting, where
the weight of each estimate depends on the length
of time since it was last updated. This approach is a
moving average of a predefined number (or proportion)
of the most recent forecasts. For example, the algorithm
might use the latest 30 updates or the most recent 20%
of updates.
(2) Heterogeneity of Individual Skills. Prediction markets
automatically provide a mechanism for incorporating
differences in forecaster knowledge and skill. In the
short run, the effect of an order depends on the number
of shares that a trader wants to buy or sell. Order size
can be construed as a measure of one’s confidence in
a position. In finite markets,
3
larger orders are more
influential. For a rational, risk-averse trader, the amount
invested in a given contract should be proportional
to the difference between the trader’s private beliefs
and the market price (Wolfers and Zitzewitz 2006). In
the long run, traders who place correct bets tend to be
rewarded with higher earnings, and their increased
wealth affords them greater power to influence future
prices. This feature of markets is central to the marginal
trader hypothesis, which stipulates that the efficiency of
prediction markets is driven by a minority of unbiased
and active participants who wield corrective influence
(Forsythe et al. 1992).
If some forecasters in prediction polls are consis-
tently better than others, aggregation algorithms can
improve performance by placing greater weight on
the predictions of forecasters who are more likely to
be accurate. We found that prior accuracy and belief
updating history were robustly indicative of future per-
formance and incorporated these into our aggregation
algorithms.
(3) Miscalibration. Markets provide incentives for
individuals to correct systematic biases, such as over-
confidence or underconfidence. A rational trader would
place bets that are profitable in expectation, realigning
prices with historical base rates. Prediction markets
generally produce adequately calibrated prices, with
the exception of the favorite long-shot bias.
Despite the well-known tendency for individuals
to be overconfident when making probability judg-
ments, simple aggregates of prediction polls tend to be
underconfident (Baron et al. 2014, Satopää et al. 2014b).
In other words, polls tend to produce average forecasts
that are too close to the ignorance prior, e.g., 50%
for a two-option question. Baron et al. (2014) discuss
two reasons for aggregate forecasts appearing to be
underconfident. First, because the probability scale is
bounded, noise in individual estimates tends to push
mean probability estimates away from the extremes.
This challenge is partially addressed by using the
median rather than the mean to aggregate predictions.
Second, although each individual may possess only
3
Much of the finance literature is concerned with infinite markets in
which individual traders have no ability to influence market prices.
We use markets with several hundred forecasters. Individual traders
can impact prices, at least in the short run.

Atanasov et al.: Prediction Markets vs. Prediction Polls
Management Science, Articles in Advance, pp. 1–16, © 2016 INFORMS 5
a portion of knowable information, the aggregate is
generally based on more information than any single
forecast. The tendency of average predictions to be bet-
ter informed but less extreme than individual estimates
makes recalibration beneficial.
2.5. Hypotheses
We offer two hypotheses about the relative accuracy
of methods. The first is driven by the importance of
outdated forecasts, inattention to past performance,
and underconfident estimates in prediction polls.
Hypothesis 1.
Prediction markets will outperform pre-
diction polls when forecasts are combined using a simple
mean.
The second hypothesis reflects the ability of statistical
algorithms to address these aggregation challenges.
Prior research also suggests that teaming confers a
consistent advantage to the accuracy of polls (Mellers
et al. 2014).
Hypothesis 2.
Prediction polls will outperform predic-
tion markets when probability estimates are elicited in teams
and combined with algorithms featuring decay, differential
weighting, and recalibration.
3. Experiment 1: Methods
The Aggregative Contingent Estimation (ACE) tour-
nament was sponsored by IARPA and lasted for four
years (from 2011 to 2015). The first year did not have
a continuous double auction prediction market. The
comparison does not cover this period. Poll data from
the first year were used to derive optimal parameters
for poll aggregation algorithms. Experiment 1 and
Experiment 2 report results from the second and third
year of the tournament, which ran from June 19, 2012
to April 10, 2013, and from August 1, 2013, to May 10,
2014, respectively.
Forecasters were recruited from professional societies’
email lists, blogs, research centers, alumni associations,
and personal connections. Entry into the tournament
required a bachelor’s degree or higher as well the
completion of psychological and political knowledge
tests, which took approximately two hours to compete.
Participants were an average of 36 years old and
mostly male (83%). Almost two-thirds (64%) had some
postgraduate training.
3.1. Experimental Design
In this study, more than 1,600 individuals participated
across 114 questions, amounting to approximately
65,000 and 54,000 forecasts from independent and
team polls, and 61,000 market orders. Participants
received $150 payment if they participated in at least
30 questions during the season.
We compared the accuracy of forecasts produced by
two types of prediction polls (teams and individuals)
Table 1 Participants Across Elicitation Methods
Prediction polls
Prediction
markets Teams Individuals
Number with at least one 535 565 595
order or one forecast
Gender (% male) 84 83 82
Age 3509 3501 3508
412035 410095 411065
Education (% with 63 65 62
advanced degrees)
Proportion with at least one 76 81 85
order or forecast (%)
Attrition rate (%) 11 14 18
and a prediction market (CDA) by randomly assigning
individuals to the three experimental conditions.
4
There
were no significant differences at baseline across the
three experimental conditions. Although the number of
forecasters assigned to prediction markets and predic-
tion polls was about 700, we focus only on forecasters
who made at least one forecast or market order. These
samples ranged from 535 to 595. Participants assigned
to prediction markets were less likely to submit at
least one prediction than those in prediction polls
(76% versus 81% and 85%, respectively). Attrition
rates were 11%, 14%, and 18% for prediction markets,
team, and independent prediction polls.
5
Table 1shows
demographic and retention data.
3.2. Elicitation Methods
In prediction markets, each contract traded at prices
between $0 and $1 and resolved at $1 if the event
occurred, $0 otherwise. Each participant started the
season with $100 in cash and received additional $225
in increments of $5 over the course of the season. Price
history was public information, as was the order book,
which displayed the six highest bids and the six lowest
ask prices. Forecasters had access to a dashboard, which
showed their portfolio holdings, including profits or
losses per question, as well as cash available to spend.
The leaderboard featured the top 50 forecasters in
terms of total net worth. The aggregate probability
forecasts assessed were based on the last price as of
midnight Pacific time.6
4
Additional conditions were investigated but are not discussed here
(see Mellers et al. 2014).
5
Sample sizes reported here are higher than those reported in Mellers
et al. (2014), because the current study used data from participants
who submitted at least one forecast or market order, whereas Mellers
et al. (2014) focused on participants who submitted enough forecasts
to evaluate their individual performance. The less restrictive inclusion
criteria in the current study resulted in higher attrition rate estimates.
6
In the first three months of the tournament, participants were
randomly assigned to four markets of approximately 125 individuals.
Then two markets of approximately 250 individuals each were
formed. Earnings and positions were carried over to the combined
markets. See the online appendix.

Atanasov et al.: Prediction Markets vs. Prediction Polls
6Management Science, Articles in Advance, pp. 1–16, © 2016 INFORMS
In team and independent prediction polls, partici-
pants provided probability forecasts. Accuracy was
measured using the Brier score, a proper scoring rule
(Brier 1950). Scores varied from 0 (best) to 2 (worst).
Forecasters in independent prediction polls worked
independently. After a question resolved, performance
was assessed as the mean daily Brier score.
7
Scores
were averaged across questions, and the 50 forecast-
ers with the lowest Brier scores were featured on a
leaderboard.
In team polls, forecasters received a 30-minute train-
ing module about how to build a high-functioning
team and hold each other accountable. Teams also
had access to collaborative tools allowing them to
write comments, use team mail to send personal mes-
sages, and receive information about the forecasts of
their teammates. Accuracy scores for team members
appeared on a within-team leaderboard. A separate
leaderboard compared performance across teams. Team
consensus was not required for submitting a forecast.
Each team member submitted their own predictions
and team scores were based on median forecaster’s
Brier score for each forecasting question.
The following analogies are useful in portraying the
information structure of markets and polls. Prediction
market traders form a mesh network, in which traders
exchange information by placing orders. A hub-and-
spokes analogy is most apt for describing independent
polls: each individual makes solo predictions but is
unable to exchange information with others. These
independent predictions are aggregated statistically.
Team-based polls resemble an archipelago, where each
team is an island: forecasters can freely communicate
within, but not across teams.
3.3. Questions and Scoring
Participants in Experiment 1 submitted probability
estimates or market orders for 114 geopolitical ques-
tions using a Web interface. Questions were released
throughout the forecasting season in small batches.
Question duration varied between 2 and 410 days, with
a median of 104 days. Forecasters were encouraged to
update their beliefs as often as they wished until the
question was resolved.
All questions had to pass the clairvoyance test: the
resolution criteria had to be specified clearly enough to
produce a definitive answer to the question. Another
desirable question property was the 10/90 rule: at time
of launch, events in question should be deemed more
than 10%, but less 90% likely.
Results are based on aggregate forecasts. Probability
forecasts and market orders were collected at the same
7
For questions with three or more ordered answer options, we
applied an ordered version of the scoring rule, which provides
partial credit (or penalty reduction) for placing high probability
estimates on answer options closer to the correct answer.
time each day. Accuracy of both methods was assessed
using the average daily Brier score, similar to the
one used for individual prediction polling forecasters,
except that aggregate forecasts were available every
day each question was open, so carrying forward old
forecasts was unnecessary.
3.4. Aggregation Algorithms in Prediction Polls
As discussed above, a critical component of poll-based
prediction is the aggregation of individual predictions
into a single forecast. The Good Judgment team used a
weighted average:
¯
pt1 k1 l =1
Pt1 i 8dt1 i1 l ×wt1 i1 l 9

×X
t1 i
8dt1 i1 l ×wt1 i1 l 9×pt1 i1 k1 l 1(1)
where
¯
pt1 k1 l
is the weighted probability aggregate across
forecasters i, that depends on time t, question l1 and
outcome k. The decay value, d
t1 i1 l
1was set to 1 for
recent forecasts, and 0 otherwise. The decay parameter
was based on 20% of total forecasters. For example, in
a sample of 500 forecasters attempting a given question,
20% decay means that the most recently updated
forecasts for 100 forecasters would be included.
The weight of a forecast, w1 depended on time t,
forecaster i, and question l. We used performance
weighting with two variables: prior accuracy and
frequency of belief updating. Weights for past accuracy
were based on an individual’s mean Brier score on
all closed questions at that time. Raw Brier scores
varied widely across questions: the average participant
scored 0.02 on the “easiest” question and 1.5 on the
“hardest.” To account for this, we standardized scores
within questions. Past accuracy weights were not set
until the first 10 questions in a season had closed.
Weights for belief-updating frequency were based on
an individual’s number of forecasts for the current
question up to this point.
The overall weight was the product of weights for
accuracy and frequency updating. Prior accuracy was
the cumulative mean standardized Brier score for all
resolved questions the forecaster attempted. Values
were rescaled to the range [0.1 to 1]; the least accurate
forecaster would receive 10% of the weight of the most
accurate forecaster. The same rescaling was applied
to the number of times a forecaster updated his or
her probability estimate. Variables were combined as
follows:
wt1 i1 l =c
t1 i ×f
t1 i1 l 1(2)
where the accuracy score, c, calculated at time tand
forecaster i1 was raised to the exponent . The fre-
quency score, f 1 calculated at time tfor forecaster ion
question l1 was raised to the exponent . Weights were

Atanasov et al.: Prediction Markets vs. Prediction Polls
Management Science, Articles in Advance, pp. 1–16, © 2016 INFORMS 7
exponents, and higher exponents produced weights
with greater variation.
Finally, we recalibrated the aggregate forecasts as
follows:
ˆ
pt1 k1 l =¯
pa
t1 k1 l
¯
pa
t1 k1 l +41−¯
pt1 k1 l 5a1(3)
where
¯
p
is the original, nontransformed probability
from the Equation (1),
ˆ
p
is the transformed probability
estimate, and ais the recalibration parameter. When
a=1, the aggregate is unchanged. At a > 1, aggregates
are more extreme, at a < 1, aggregates are less extreme.
All parameters were set sequentially, starting with
temporal decay, prior performance weighting, belief-
updating weighting, and ending with recalibration,
with the goal of minimizing error in the resulting
aggregated forecasts. At each step, we used elastic net
regularization (Zou and Hastie 2005) to minimize the
overfitting. This method penalizes entry into the model
and the magnitude of each parameter therein, which
increases the chances that the parameters obtained at
each step capture global versus local optimal weights.
4. Experiment 1: Results
4.1. Relative Accuracy
We assessed the accuracy of methods using paired
t-tests with questions as the unit of analysis. Predic-
tion markets significantly outperformed the simple
means of prediction polls. Mean Brier scores for team
and independent polls were 0.274 and 0.307, respec-
tively, significantly higher than the prediction market’s
mean Brier score of 0.214 (paired t41135=4045 and
t41135=5092, p < 00001 for each). Relative to team and
independent polls, prediction markets were 22% and
30% more accurate. The results, shown in Figure 1,
support Hypothesis 1.
4.1.1. Aggregation Parameters. Hypothesis 2states
that prediction poll forecasts that are combined statisti-
cally to reflect temporal decay, differential weighting,
and recalibration are more accurate than prediction
markets. To test this hypothesis, we estimated the
parameters of the aggregation rule by fitting the algo-
rithm to an independent set of year 1 data and applying
it out of sample to year 2 data.
Figure 1shows how each component of the aggrega-
tion rule decreased errors relative to the mean. With all
three components, team prediction polls were signifi-
cantly more accurate than the prediction market, yield-
ing Brier scores that were 13% lower (mean difference =
0.027, paired t41135=2069, p=00008). This result sup-
ports Hypothesis 2. Forecasts from independent polls
that were aggregated statistically outperformed predic-
tion market prices by 9%, but the differences did not
reach statistical significance (t41135=1051, p=00136).
Figure 1 Mean Brier Scores for Independent Prediction Polls, Team
Prediction Polls, and Prediction Markets with 114 Questions
0
0.100
0.200
0.300
0.400
Simple
mean
Decay
only
Weighting
only
Extremizing
only
All three
components
Brier score
Independent prediction polls
Team prediction polls
Prediction markets
Note. Errors bars denote one standard error of the difference between each
polling method and the prediction market.
To summarize, results supported Hypothesis 1: pre-
diction markets were more accurate than prediction
polls when poll forecasts were combined using a simple
mean. In addition, Hypothesis 2was partially sup-
ported. Forecasts from team prediction polls combined
using the full algorithm outperformed the prediction
market. Furthermore, the accuracy of independent
prediction polls and prediction markets was not reli-
ably different. Prediction polls outperformed predic-
tion markets when forecasters share information and
aggregation was conducted with attention to timing,
engagement, previous accuracy, and extremeness.
Next, we examine the stability and generalizability of
the results by examining effects of (a) parameters used
in the aggregation of forecasts from prediction polls,
(b) different scoring rules, and (c) different measures
of accuracy. Finally, we assess how relative accuracy
varies with question duration and the number of active
forecasters across questions.
4.2. Sensitivity to Aggregation Parameters
The better performance of prediction polls was based
on parameters derived from a prior tournament year
(e.g., out-of-sample data), but it is possible that slight
changes in parameter settings would have led to differ-
ent results. We focus on two parameters: decay and
recalibration. Figure 2(a) shows the effects of temporal
decay parameters. Using a percentage of the most
recent forecasts (20%), we obtain more accurate aggre-
gates, and the exact proportion had a limited influence
on the results.
Figure 2(b) shows effects of recalibration parame-
ters. No recalibration, (denoted as a=1) resulted in
lower accuracy relative to recalibration parameters that
pushed the aggregate forecasts toward the extremes
(a > 1). However, a wide range of values for apro-
duced comparable levels of accuracy. For example,

Atanasov et al.: Prediction Markets vs. Prediction Polls
8Management Science, Articles in Advance, pp. 1–16, © 2016 INFORMS
Figure 2 Aggregate Performance for Independent and Team-Based Prediction Polls, Varying Temporal Decay (a) and Recalibration (b) Parameters
0.18
0.19
0.20
0.21
(a)
0
Mean Brier score
Proportion of recent forecasts retained
0.2 0.4 0.6 0.8 1.0
Mean Brier score
0.18
0.20
0.22
0.24
1.00
Recalibration constant a
1.25 1.50 1.75 2.00 2.25 2.50
(b)
Polls, team
Polls, independent
Note. All other parameters remain at optimized levels.
the team recalibration parameter was a=105. Results
would have improved if a=2, but the improvement in
accuracy would have been less than 2%. To summarize,
within limits, changes in both the parameters had
limited impact on the accuracy of prediction polls.
The advantage of team over independent polls was
not a statistical artifact of our selection of aggregation
parameters for the team forecasts.
If statistical aggregation improved the accuracy
of prediction polls, should statistical aggregation be
applied to data from prediction markets? If market
prices suffered from excess volatility, accuracy might
increase with an exponential smoothing over the end-
of-day prices. A decay function with an exponent of 0
means that old prices get no weight, which is the same
as using the last price, whereas a decay parameter of 1
means that the last five end-of-day prices are equally
weighted. As shown in Figure 3(a), placing no weight
on past prices achieved the highest accuracy.
Several studies have demonstrated that prediction
markets suffer from the favorite long-shot bias (Page
and Clemen 2012, Snowberg and Wolfers 2010). Roth-
schild (2009) showed that when market prices are
recalibrated (pushed toward the extremes), accuracy
improved. We applied the recalibration to market prices
(a=105). As we can see in Figure 3(b), extremizing
market prices made accuracy worse, increasing Brier
scores. Recalibration of prices away from the extremes
also worsened accuracy. In summary, the last price
yielded the most accurate probability forecast in the
prediction markets.8
8
We examined different ways of combining estimates from the two
prediction markets and found the simple mean of latest market
prices to be effective. For example, we calculated bid-ask spreads for
the two parallel markets and constructed a weighted mean price,
placing more weight on the market with the lower bid-ask spread
for each question-day combination but this weighting scheme did
not improve accuracy. See the online appendix.
4.3. Alternative Scoring Rules
Do prediction markets outperform team prediction
polls under different scoring rules? We examine three
additional rules: logarithmic, spherical, and absolute
distance. When a method assigns a value of f
c
to the
correct answer of a binary question, these rules are
defined as follows:
Brier Scoring Rule =2×61−fc721
from 2 (worst) to 0 (best)
Logarithmic Scoring Rule =ln4fc51
from −  (worst) to 0 (best)
Spherical Scoring Rule =fc/6f 2
c+41−fc5271/21
from 0 (worst) to 1 (best)
Absolute Distance Rule =2× 1−fc1
from 2 (worst) to 0 (best).
All rules except the absolute distance rule are strictly
proper, meaning that forecasters’ optimal strategy
is to report their true beliefs. Improvement of team
prediction polls over prediction markets was 11%
with the logarithmic rule, 2% with the spherical rule,
and 6% with the absolute distance rule. Forecasts
from team prediction polls were more accurate than
prediction market prices with all scoring rules (p < 0005).
The independent prediction poll outperformed the
prediction market with the absolute distance rule only.
See Table 2.
4.4. Calibration, Discrimination, and AUC
Brier scores can be decomposed into three components:
variability, calibration, and discrimination (Murphy
1973). Variability is independent of skill and simply
reflects the base rate for events in the environment, so
it will not be discussed further. Calibration refers to
the ability to make subjective forecasts that, in the long
run, coincide with the objective base rates of events.

Atanasov et al.: Prediction Markets vs. Prediction Polls
Management Science, Articles in Advance, pp. 1–16, © 2016 INFORMS 9
Figure 3 Performance for Prediction Markets for Varying Temporal Decay (a) and Recalibration (b) Parameters
0.20
0.21
0.22
0.23
0.24
0
Mean Brier score
Decay constant
0.2 0.4 0.6 0.8 1.0
0.20
0.21
0.22
0.23
0.24
Mean Brier score
1.00
Recalibration constant a
1.25 1.50 1.75 2.00
(a) (b)
Table 2 Relative Performance as Judged by Four Scoring Rules
Statistical aggregation
Prediction Team Independent
Scoring rule market prediction poll prediction poll
Brier 0021 (0.31) 0019 (0.27)∗∗ 0019 (0.31)
Log −0034 (0.41) −0030 (0.34)∗∗ −0031 (0.41)
Spherical 0088 (0.18) 0090 (0.16)∗0089 (0.18)
Absolute distance 0045 (0.40) 0042 (0.36)∗0040 (0.39)∗∗
Note. Means (SD) across 114 questions.
∗p < 0005; ∗∗p < 0001; paired samples t-test.
Calibration error would be zero if forecasts exactly
matched event base rates. Probability predictions are
said to be “well calibrated” if average confidence and
the percentage of correct forecasts are equal. Before
recalibration, team and independent prediction polls
had calibration errors of 0.029 and 0.034 because of
underconfidence. The prediction market produced
lower calibration error, 0.009. Recalibration corrected
the underconfidence in aggregate poll forecasts and
improved the performance of prediction polls. This
was evident from the decrease in Brier scores resulting
from recalibration, as shown in Table 3. Figure 4shows
how recalibration decreased calibration errors for team
and independent prediction polls.9
Discrimination refers to the tendency to place high-
probability predictions on events that occur, and low
estimates on those that do not. The discrimination
values for statistically aggregated prediction polls were
higher than those for the prediction market (see Table 3).
Discrimination was better for both independent and
team-based prediction polls with recalibration, relative
to markets.
Area under the receiver operating characteristic curve
(A-ROC) is a nonparametric measure of discrimination
9
The underconfidence of aggregate team poll of forecasts is not
relevant to debates about group polarization, groupthink or related
phenomena. It is mathematically possible that each individual team
produces polarized, overconfident predictions, but averaging across
teams produces forecasts that are underconfident.
Table 3 Calibration and Discrimination for Statistically Aggregated
Polls and Markets Based on Brier Score Decomposition
Prediction Team Independent
Method market prediction poll prediction poll
Calibration 00009 00010 00008
Discrimination 00355 00384 00375
Variability 00585 00585 00585
Note. Recalibration based on year 1 was applied to year 2 prediction poll
forecasts.
derived from signal detection theory (Swets 1996). The
ROC curve plots the probability of a hit (true positive)
against that of a false alarm (false positive). Perfect
resolution means all of the probability mass is under
the curve. ROC curves for prediction markets, as well
as algorithmically aggregated independent and team
polls are shown in Figure 5, grouped by early, middle,
and late thirds of each question. In the first third of the
questions, prediction markets yielded worse ROC scores
than prediction polls. Markets and prediction polls
were tied in the middle and late stages. The resolution
advantage of team and independent prediction polls
was stronger at the start of the questions, when there
was maximum uncertainty about the outcome.
In summary, team and independent prediction polls
either outperformed or tied prediction markets using
scoring rules (logarithmic, spherical score, and absolute
distance), alternative accuracy measures (discrimination
and area under the ROC), and separate periods of the
forecasting window (early, middle, and late). Next, we
ask whether team prediction polls’ performance edge
varies with properties of the forecasting questions.
4.5. Duration of Forecasting Questions10
Does the relative performance of markets and polls
depend upon the duration of the forecasting question?
10
We also examined other potential moderators, such as question-
level uncertainty ratings, close call index ratings, status quo versus
change outcome classification, question type (binary, conditional, and
multioption). None of these variables were significant moderators of
relative performance.

Atanasov et al.: Prediction Markets vs. Prediction Polls
10 Management Science, Articles in Advance, pp. 1–16, © 2016 INFORMS
Figure 4 Calibration Plots for Independent Prediction Polls, Before (a) and After (b) Recalibration; Team Prediction Poll Forecasts, Before (c) and
After (d) recalibration; Prediction Markets, Before (e) and After (f) Recalibration (a=105)
21% 21%
17%
12%
9%
6%
6% 5%
2%
2%
0
0.25
0.50
0.75
1.00
Observed frequency
0
0.25
0.50
0.75
1.00
Observed frequency
49%
12% 8%
6%
4%
4%
2%
3%
5%
6%
(a) Independent prediction polls—no recalibration (b) Independent prediction polls—recalibration
26% 21%
16%
10%
7%
5%
6%
6% 2%
2%
45%
15%
8%
7%
4%
4%
3% 4%
6% 5%
48%
(c) Team prediction polls—no recalibration (d) Team prediction polls—recalibration
(e) Prediction markets—no recalibration (f) Prediction markets—recalibration
0
Mean forecast
0.25 0.50 0.75 1.00
0
Mean forecast
0.25 0.50 0.75 1.00 0
Mean forecast
10%
3%
3%
3%
2%
4%
3%
5%
9%
60%
0.25 0.50 0.75 1.00
0
Mean forecast
0.25 0.50 0.75 1.00
0
Mean forecast
0.25 0.50 0.75 1.00
0
Mean forecast
0.25 0.50 0.75 1.00
0
0.25
0.50
0.75
1.00
Observed frequency
0
0.25
0.50
0.75
1.00
Observed frequency
0
0.25
0.50
0.75
1.00
Observed frequency
0
0.25
0.50
0.75
1.00
Observed frequency
6%
4%
3%
3%
4%
5%
5%
8%
13%
Notes. Horizontal axis is divided in 10 bins, each spanning 10% on the probability scale. Percentages in chart denote the proportion of forecasts that fall in
each bin.

Atanasov et al.: Prediction Markets vs. Prediction Polls
Management Science, Articles in Advance, pp. 1–16, © 2016 INFORMS 11
Figure 5 (Color online) ROC Curves for Forecasts From First Third, Middle Third, and Last Third of Days When Each Question Was Open
(Shown with Statistically Aggregated Polls and Prediction Markets)
Last third
0
0.25
0.50
0.75
1.00
0 0.25 0.50 0.75 1.00 0 0.25 0.50 0.75 1.00 0 0.25 0.50 0.75 1.00
False positive rate
True positive rate
Polls, independent Polls, team Prediction market
First third Middle third
The favorite long-shot bias in prediction markets has
been viewed as a rational response to liquidity con-
straints that occurs when questions are open for long
periods of time. If payoffs do not materialize in a timely
fashion, individuals may not wish to place expen-
sive long-term bets on events, especially on favorites,
because investments have opportunity costs (Page and
Clemen 2012).
To examine whether prediction polls outperformed
prediction markets on shorter and longer questions, we
performed a median split based on the number of days
questions were open (median duration was 105 days).
We calculated daily Brier scores for each method on
each question, averaged scores across questions, and
rescaled question duration to the 0%–100% scale, where
0% denotes the first day and 100% denotes the last
day of a question. Figure 6shows the results of this
exercise.
We used a mixed-effects linear model to determine
whether daily Brier scores of prediction markets and
team prediction polls varied by question duration,
accounting for clustering of errors within questions
and for the number of days until the question was
resolved. We detected a significant interaction (t=4032,
p < 00001) between method (market versus team poll)
and question duration (shorter versus longer questions).
Prediction markets performed on par with team polls
on shorter questions, but produced larger errors on
longer questions. This pattern, evident in Figure 6, is
consistent with Page and Clemen’s (2012) discussion of
the favorite long-shot bias, except the current results
concern overall accuracy rather than calibration.
4.6. Forecasting Activity
Better performance in team prediction polls might
have been due to greater forecaster activity relative
to prediction markets. We examine two measures of
activity—number of questions attempted, and engage-
ment within a question. Engagement was defined
differently across methods. In prediction polls, we used
frequency of updating, and in the prediction market,
we used number of orders.
As shown in Table 4, forecasters in the prediction
market attempted fewer questions than those in predic-
tion polls. Averages were 34, 52, and 66 in the market,
the team poll, and independent poll, respectively. How-
ever, the average number of orders placed per question
in the market was larger than the average number of
forecasts made (over persons and questions) in the
polls.11
Mellers et al. (2015a) found that, in prediction
polls, frequency of updating was a better predictor
of accuracy than the number of forecasting questions
attempted. We replicated these results. In team and
independent prediction polls, belief updating correlated
with average standardized Brier scores (r= −0018,
t41
1
1585=6012, p < 00001), whereas number of questions
attempted was not significant. A similar pattern held
in prediction markets. The average number of orders
placed per question was the best behavioral predictor
of total earnings (r=0055, t45345=20005, p < 00001),
whereas the number of questions attempted was a
weaker predictor of earnings (r=0020, t45345=4080,
p < 00001).
The focus on mean activity levels belies an important
distinction between markets and polls. Prediction mar-
ket orders were skewed across market participants, as
evidenced by the higher standard deviation for orders
per question. More than 50% of all orders were placed
by the most active 5% of prediction market participants.
In contrast, fewer than 30% of all forecasts were made
by the most active 5% of prediction poll participants.
Simple forecast counts offer an incomplete picture of
relative influence, due to the differential weighting of
11
Multiple orders or forecasts on a question on a given day only
counted as one order or forecast.

Atanasov et al.: Prediction Markets vs. Prediction Polls
12 Management Science, Articles in Advance, pp. 1–16, © 2016 INFORMS
Figure 6 (Color online) Performance over Time for Statistically Aggregated Polls and Prediction Markets When Questions Lasted Less (Left) and
More (Right) Than the Median, 105 Days
Short duration Long duration
0
0.1
0.2
0.3
0 0.25 0.50 0.75 1.00 0 0.25 0.50 0.75 1.00
Percentage of time since start
Mean daily brier score
Method
Polls, independent Polls, team Prediction market
Note. Local regression (Loess) smoothed curves are added to illustrate differences over time.
individuals in polls and the varying transaction sizes
in markets. To add a more refined perspective, we
calculated the relative weight of each forecaster on
each question across all days the question was open,
accounting for decay, accuracy, and updating weights.
We then summed the weights across questions to obtain
a measure of overall influence per forecaster throughout
the season. The 20 highest-weighted forecasters in their
condition accounted for 27% and 34% of the weights
for team and independent polls, respectively.
Table 4 Activity Patterns Across Elicitation Methods
Team Independent
Prediction prediction prediction
Activity metric market poll poll
Number of questions attempted 34 52 66
(26) (42) (45)
Number of orders/forecasts per 1.9 1.8 1.6
forecaster per question (2.9) (1.5) (1.8)
Number of orders/forecasts per 63 88 102
forecaster per year (106) (107) (134)
Number of forecasters per 136 228 299
question (66) (85) (85)
Number of orders/forecasts 260 473 491
per question (201) (218) (221)
Note. Cell entries are means; standard deviations shown in parentheses.
We performed a parallel calculation in prediction
markets, except we used total transaction values as a
marker of influence. For example, a “buy” order for
three shares at $0.40 would have a total cost of $1.20,
and the matching sell order will cost $0.60 per share
or $1.80 in total. We summed up transaction values
for each person within a question, then calculated the
proportion of all dollars spent by each participant across
all questions. The top 20 traders accounted for 33% of
the overall transaction values. To the extent that weights
and transaction values can be compared as measures
of relative influence across platforms, we can say
that highly active forecasters tended to exercise more
influence in markets than in team polls. Independent
polls and prediction markets were generally similar.
The last row of Table 4shows that, on average, more
forecasts/orders were made per question in prediction
polls than in prediction markets. Is the greater accuracy
of team prediction polls driven by larger sample sizes?
In the independent prediction polls conditions, we
can estimate the impact of sample size by randomly
selecting forecasters and removing them from the
aggregate. These analyses showed that the market and
poll performance was similarly sensitive to smaller
samples. See the online appendix.
We now proceed to Experiment 2, which took place
in the year following Experiment 1. This study provides

Atanasov et al.: Prediction Markets vs. Prediction Polls
Management Science, Articles in Advance, pp. 1–16, © 2016 INFORMS 13
a robustness check for Experiment 1 by applying the
similar elicitation and aggregation methods to a new
set of questions.
5. Experiment 2: Methods
More than 1,900 individuals participated between
August 1, 2013, and May 10, 2014. One hundred and
forty seven questions were asked. Participants received
a $250 payment conditional on forecasting in at least
30 questions over the course of the season.
Experiment 2 is based on data from the third year of
the ACE tournament, directly following Experiment 1,
which covered the second year. The experimental
design in Experiment 2 followed the Experiment 1 con-
vention: new forecasters were randomly assigned to the
three experimental conditions; continuing forecasters
remained in their original conditions (e.g., prediction
market forecasters were reassigned to a market).
Experiment 2 differed from Experiment 1 in five ways.
First, the number of participants assigned to teams
prediction polls was larger than the number of fore-
casters in independent prediction polls (N=839 and
550 in polls versus prediction markets with N=539).
Second, all market participants traded in one larger
prediction market, rather than two smaller markets, as
was the case in Experiment 1. Third, all team partici-
pants received forecasting training, whereas half of
independent poll and prediction market participants
received training. Fourth, half of the teams had a team
coach to help facilitate discussion, whereas none of
the independent forecasters and market forecasters
had a coach. Coaching had no impact on accuracy.
Fifth, only U.S. citizens were allowed in the prediction
market, whereas approximately 22% of team and 24%
of independent poll participants were non-U.S. citizens.
To address these experimental design challenges,
we excluded all non-U.S. citizens from independent
and team prediction polls. After this, we were left
with 456 and 605 independent and team participants,
respectively. The sample included 44,000 predictions in
independent polls, 61,000 predictions in team polls,
and 71,000 orders in prediction markets. Of course,
exclusion of non-U.S. citizens from teams was not a
Table 5 Forecasting Accuracy for Experiment 2 Using Four Scoring
Rules
Statistical aggregation
Prediction Team Independent
Scoring rule market prediction poll prediction poll
Brier 0024 (0.33) 0018 (0.29)∗∗ 0025 (0.33)
Log −0040 (0.47) −0030 (0.44)∗∗ −0041 (0.53)
Spherical 0086 (0.19) 0090 (0.16)∗∗ 0086 (0.19)
Absolute distance 0049 (0.40) 0036 (0.37)∗∗ 0043 (0.42)∗
Note. Means (SD) shown for 147 questions.
∗p < 0005; ∗∗p < 0001; paired samples t-test.
complete solution. Although removing team members’
probability was straightforward, we could not eliminate
the information they shared with their U.S. teammates.
The proportion of non-U.S. citizens in each team ranged
from 9% to 42%, with a mean of 24%. There was
no correlation between those proportions and team
accuracy. Thus, there was no correlational evidence
having more non-U.S. citizens in a team was associated
with higher accuracy, but such a relationship cannot be
ruled out. Despite these nonexperimental assignment
features, we believe Experiment 2 provides additional
evidence about the relative performance of methods.
6. Experiment 2: Results
The mean Brier score across all 147 questions for
the prediction market was 0.24 (SD =0.34). Simple
means of team polls and independent polls were 0.22
(SD =0.22) and 0.31 (SD =0.20), respectively. The
prediction market was more accurate than the simple
mean of independent poll forecasts (paired t41465=3048,
p < 0001), but no different from the simple mean of team
poll forecasts. Thus, results were partially consistent
with Hypothesis 1; prediction markets outperformed
simple means of independent polls, but not simple
means of team polls.
When we added temporal decay, differential weight-
ing, and recalibration to the aggregation algorithm,
team prediction polls significantly outperformed pre-
diction markets for all four scoring rules, shown in
Table 5. Independent polls were tied with markets,
except for polls’ significant outperformance according
to the absolute distance rule.
7. Discussion and Conclusions
In Experiment 1, a randomized experiment comparing
prediction markets versus competitive, dynamic predic-
tion polls, simple aggregations of prediction polls were
less accurate than market prices. The prediction market
performed well with no additional benefit from statisti-
cal aggregation. The last price of the market was the
single most accurate estimate that could be extracted
from market data. Statistical aggregation improved
accuracy in prediction polls, helping team polls to
significantly outperform markets. This advantage held
up across different scoring rules and accuracy mea-
sures. Team prediction polls were better than prediction
markets on longer questions and at the early stages
of the questions, the periods of greatest uncertainty.
Prediction markets were more accurate than polls at
the start of short-duration questions. The advantage of
team polls over markets extended to Experiment 2.
7.1. What Gives Prediction Polls Their Edge?
The stronger performance of prediction polls can be
attributed to several factors. First, to perform well in a

Atanasov et al.: Prediction Markets vs. Prediction Polls
14 Management Science, Articles in Advance, pp. 1–16, © 2016 INFORMS
prediction poll, one needs substantive knowledge and
skill in converting insights into probability judgments.
In prediction markets, one needs both timely, relevant
information, and an understanding of the strategic
aspects of market trading. The need for strategic sophis-
tication in markets may have served as a barrier to high
performance for otherwise knowledgeable
forecasters
.
Second, in settings where questions are posed and
answered sequentially, information about prior activity
and performance can be used to amplify the voices
of the most careful and accurate forecasters in pre-
diction polls and dampen the clamor of inaccurate
ones. Prediction polls can use differential weighting
based on prior performance. In prediction markets,
individuals were weighted based on order size in the
short run and accumulated wealth in the long run. Our
results suggest that the weighting mechanisms used
in prediction polls are at least as effective as those
inherent to markets.
Third, crowd aggregates of probability judgments
are usually underconfident, even when individual
forecasts are not (Ranjan and Gneiting 2008). When
such patterns persist over time, recalibration tends
to improve aggregate accuracy. This problem was
effectively addressed with the recalibration formula we
used in both Experiment 1 and Experiment 2.
Fourth, the incentives for sharing information are
better in team prediction polls than in markets. Contin-
uous double auction markets are a zero sum game: one
trader’s gain is another’s loss. By contrast, team predic-
tion polls are structured so that people have incentives
to discuss their beliefs and rationales with team mem-
bers, even if they think someone else knows more.
Teams are not limited to prediction polls; intrateam
collaboration could, in principle, be incorporated into a
prediction market.
Fifth, the advantage of polls is especially great when
the number of active forecasters is small. A prediction
poll of one person may be useful, but there is no
market with a single trader. Even with a handful of
people, the simple polling averages provide reasonable
results (Mannes et al. 2014), while double auction
prediction markets suffer from liquidity problems.
Scoring rule markets (e.g., LMSR) are designed to
solve the small market problem, but they require
some interventions from tournament designers to set
a liquidity parameter. Scoring rule markets are more
similar to polls in that they do not operate as a zero-
sum game. These markets may be more conducive to
information sharing among forecasters. For settings
with large numbers of forecasters per question, Hanson
(2003) suggests that prediction polls may “suffer from
a thick market problem, namely, how to produce a
single consensus estimate when different people give
differing estimates.” (p. 108). We show that appropriate
statistical aggregations can solve this problem.
In sum, team prediction polls created a mix of
intrateam cooperation and interteam competition. This
mixed environment may be better than a purely com-
petitive one if individuals share information with each
other and pool their knowledge. Team prediction polls
may have a cognitive advantage via information shar-
ing. Team prediction polls may have a motivational
advantage if team members encourage each other to
update their forecasts regularly. More complex forms
of both cooperation and competition may increase
accuracy relative to the purely competitive market
environment. Ostrovsky (2012) discusses information
pooling challenges in markets populated by strategic,
partially informed traders. These challenges can be
addressed by introducing incentives for cooperation,
such as those used in team-based polls. In addition,
teaming has been shown to improve problem solving in
probability and reasoning tasks, including in auctions
(Maciejovsky et al. 2013).
7.2. Trade-offs
There are pros and cons to prediction polls and predic-
tion markets that go beyond pure accuracy. Prediction
markets aggregate forecasts instantly and automatically.
The instant aggregation and price availability to all
market participants is a great way to promote trans-
parency. Markets provide a decentralized method for
information discovery. On the other hand, prediction
polls perform well even when all participants pro-
vide independent estimates, with no crowd feedback,
and are aggregated statistically. This feature may be
especially useful in settings where managers need to
collect predictions but have reasons to avoid sharing
aggregate forecasts with participants.
Prediction polls require a statistical aggregation rule
to more effectively pool multiple opinions. The need
for devising statistical aggregation rules increases the
responsibilities of the manager running the prediction
polls. However, the components of effective statistical
algorithms are fairly general, and parameter values
for temporal decay and recalibration are relatively
insensitive to misspecification. Individual weighting of
forecasters is done after tournament designers achieve
sufficient knowledge about participants. Based on our
experience, “seeding” polls with 20–25 initial ques-
tions should be sufficient to derive stable aggregation
parameters. Aggregation could easily be built into
prediction polling software to provide a more seamless,
off-the-shelf application.
7.3. Future Directions
Although we present the first long-term, experimental
comparison of prediction polls and prediction markets,
our results are by no means the final word. The tour-
nament paid participants for participation rather than
accuracy. If payments had been based on performance,

Atanasov et al.: Prediction Markets vs. Prediction Polls
Management Science, Articles in Advance, pp. 1–16, © 2016 INFORMS 15
the results could have differed. Studies of prediction
markets have examined the effects of real performance-
based payments relative to hypothetical-based pay-
ments, and the results are mixed. Servan-Schreiber et al.
(2004) found no effects of real monetary incentives in
prediction markets, but Rosenbloom and Notz (2006)
showed that real financial incentives produced better
forecasts. Camerer and Hogarth (1999) argued that
real monetary incentives tend to decrease response
variance. Finally, Hertwig and Ortmann (2001) found
that performance incentives tend to induce decisions
that more closely approximate rationality. Thus it is
reasonable to expect that accuracy in both prediction
markets and prediction polls with proper scoring rules
will improve with the introduction of monetary incen-
tives. However, it is less clear if the accuracy benefits of
monetary incentives will be greater in markets or in
polls.
One controversial feature of our design was the use
of a leaderboard that displayed not only absolute earn-
ings and proper scores, but also ranks. The distinction
between absolute scores and ranks has been discussed
in the expert elicitation and macroeconomic forecasting
literature. Laster et al. (1999) discuss a model, and
supporting evidence, that awards top forecasters for
deliberately deviating from the consensus. Lichtendahl
and Winkler (2007) argue that rank-based incentives
push probability forecasters toward the extremes, rel-
ative to a setting with proper scoring rules without
rankings. Although risk seeking tends to push polling
averages toward the extremes, we note that it has the
opposite effect in markets: risk-loving traders prefer
long-shot bets, which tends to push prices away from
the extremes.
The predicted patterns of overconfidence in polls
and underconfidence in markets are inconsistent with
our empirical results. Prediction poll forecasts are
slightly overconfident at the individual level (Moore
et al. 2015), but underconfident at the aggregate level,
as we show above. Prediction market prices show
neither systematic underconfidence nor overconfidence.
Prices did not exhibit the typical favorite long-shot
bias, implying that ranking system did not lead to
excessive risk taking. The lack of distortions may be
reassuring to companies that run prediction markets
and rely on leaderboards and reputational incentives
(Cowgill and Zitzewitz 2015).
Another limitation to the generality of our results is
that we tested only continuous double auction mar-
kets. We cannot generalize our results to other market
structures, such as pari-mutuel or logarithmic market
scoring rule (LMSR) markets. Future research should
examine these possibilities.
7.4. Conclusions
We show that crowds of several hundred individuals
can produce highly accurate predictions on a wide
range of political and economic topics. A comparison of
two crowd-based forecasting approaches demonstrated
that prediction polls that used teams and statistical
aggregation were more accurate than prediction mar-
kets. Our results show that it is possible to take the
elements that make prediction markets work well, such
as incentivized belief elicitation, effective aggregation,
and performance weighting, and combine these ele-
ments to a crowdsourcing method that produces even
more accurate aggregate beliefs.
Supplemental Material
Supplemental material to this paper is available at http://dx
.doi.org/10.1287/mnsc.2015.2374.
Acknowledgments
The authors thank INFORMS, Association for Psychological
Science, and Judgment and Decision Making conference partic-
ipants for helpful comments on previous drafts. This research
was supported by the Intelligence Advanced Research Projects
Activity (IARPA) via the Department of Interior National
Business Center (DoI/NBC) [Contract D11PC20061]. The
U.S. Government is authorized to reproduce and distribute
reprints for Government purposes notwithstanding any
copyright annotation thereon. The views and conclusions
expressed herein are those of the authors and should not be
interpreted as necessarily representing the official policies
or endorsements, either expressed or implied, of IARPA,
DoI/NBC, or the U.S. Government.
References
Armstrong S (2001) Combining forecasts. Armstrong JS, ed. Principles
of Forecasting (Springer, New York), 417–439.
Atanasov P, Rescober P, Servan-Schreiber E, Mellers B, Tetlock
P, Ungar L (2013) The marketcast method for aggregating
prediction market forecasts. Proc. 2013 Internat. Conf. Social
Comput., Behavioral-Cultural Modeling, and Prediction (Springer,
New York).
Baron J, Mellers BA, Tetlock PE, Stone E, Ungar LH (2014) Two
reasons to make aggregated probability forecasts more extreme.
Decision Anal. 11(2):133–145.
Berg J, Forsythe R, Nelson F, Rietz T (2001) Results from a dozen years
of election futures markets research. Plott CR, Smith VL, eds.
Handbook of Experimental Economics Results, Vol. 1 (North-Holland,
Amsterdam), 742–751.
Brier GW (1950) Verification of forecasts expressed in terms of
probability. Monthly Weather Rev. 78(1):1–3.
Camerer CF, Hogarth RM (1999) The effects of financial incentives in
experiments: A review and capital-labor-production framework.
J. Risk Uncertainty 19(1–3):7–42.
Chen Y, Chu C, Mullen T, Pennock D (2005) Information markets
vs. opinion pools: An empirical comparison. Proc. 6th ACM
Conf. Electronic Commerce (Association for Computing Machinery,
New York), 58–67.
Cowgill B, Zitzewitz E (2015) Corporate prediction markets: Evi-
dence from Google, Ford, and Firm X. Rev. Econom. Stud. 82(4):
1309–1341.
Fama EF (1970) Efficient capital markets: A review of theory and
empirical work. J. Finance 25(2):383–417.
Forsythe R, Nelson F, Neumann GR, Wright J (1992) Anatomy
of an experimental political stock market. Amer. Econom. Rev.
82(5):1142–1161.

Atanasov et al.: Prediction Markets vs. Prediction Polls
16 Management Science, Articles in Advance, pp. 1–16, © 2016 INFORMS
Goel S, Reeves D, Watts D, Pennock D (2010) Prediction without
markets. Proc. 11th ACM Conf. Electronic Commerce (Association
for Computing Machinery, New York), 357–366.
Graefe A, Armstrong J (2011) Comparing face-to-face meetings,
nominal groups, Delphi and prediction markets on an estimation
task. Internat. J. Forecasting 27(1):183–195.
Gürkaynak RS, Wolfers J (2006) Macroeconomic derivatives: An
initial analysis of market-based macro forecasts, uncertainty, and
risk. NBER Working Paper 11929, National Bureau of Economic
Research, Cambridge, MA.
Hanson R (2003) Combinatorial information market design. Inform.
Systems Frontiers 5(1):107–119.
Hayek F (1945) The use of knowledge in society. Amer. Econom. Rev.
35(4):519–530.
Hertwig R, Ortmann A (2001) Experimental practices in economics:
A methodological challenge for psychologists? Behav. Brain Sci.
24(3):383–403.
Lichtendahl KC Jr, Winkler RL (2007) Probability elicitation, scoring
rules, and competition among forecasters. Management Sci.
53(11):1745–1755.
Laster D, Bennett P, Geoum IS (1999) Rational bias in macroeconomic
forecasts. Quart. J. Econom. 114(1):293–318.
Maciejovsky B, Sutter M, Budescu D, Bernau P (2013) Teams make
you smarter: How exposure to teams improves individual
decisions in probability and reasoning tasks. Management Sci.
59(6):1255–1270.
Mannes AE, Soll JB, Larrick RP (2014) The wisdom of select crowds.
J. Personality Soc. Psych. 107(2):276–299.
Manski C (2006) Interpreting the predictions of prediction markets.
Econom. Lett. 91(3):425–429.
Mellers B, Ungar L, Baron J, Ramos J, Gurcay B, Fincher K, Tetlock P
(2014) Psychological strategies for winning a geopolitical fore-
casting tournament. Psychol. Sci. 25(5):1106–1115.
Mellers B, Stone E, Atanasov P, Rohrbaugh N, Metz SE, Ungar L,
Bishop MM, Horowitz M, Merkle E, Tetlock P (2015a) The psy-
chology of intelligence analysis: Drivers of prediction accuracy
in world politics. J. Experiment. Psychol.: Appl. 21(1):1–14.
Mellers B, Stone E, Murray T, Minster A, Rohrbaugh N, Bishop M,
Chen E, et al. (2015b) Identifying and cultivating superforecasters
as a method of improving probabilistic predictions. Perspect.
Psychol. Sci. 10(3):267–281.
Moore DA, Swift SA, Minster A, Mellers B, Ungar L, Tetlock P,
Yang H, Tenney R (2015) Confidence calibration in a multiyear
geopolitical forecasting competition. Working paper, University
of California, Berkeley, Berkeley.
Murphy AH (1973) A new vector partition of the probability score.
J. Appl. Meteorology 12(4):595–600.
Ostrovsky M (2012) Information aggregation in dynamic markets
with strategic traders. Econometrica 80(6):2595–2647.
Page L, Clemen RT (2012) Do prediction markets produce well-
calibrated probability forecasts? Econom. J. 123(568):491–513.
Pennock DM, Lawrence S, Nielsen FA, Giles CL (2001) Extract-
ing collective probabilistic forecasts from Web games. Proc.
7th ACM SIGKDD Internat. Conf. Knowledge Discovery Data
Mining (Association for Computing Machinery, New York),
174–183.
Plott C, Chen K (2002) Information aggregation mechanisms:
Concept, design and implementation for a sales forecasting
problem. Working paper, California Institute of Technology,
Pasadena.
Ranjan R, Gneiting T (2008) Combining probability forecasts. Techni-
cal Report 543, University of Washington, Seattle.
Rieg R, Schoder R (2010) Forecasting accuracy: Comparing prediction
markets and surveys—An experimental study. J. Prediction
Markets 4(3):1–19.
Rosenbloom ES, Notz W (2006) Statistical tests of real-money versus
play-money prediction markets. Electronic Markets 16(1):63–69.
Rothschild D (2009) Forecasting elections: Comparing prediction
markets, polls, and their biases. Public Opinion Quart. 73(5):
895–916.
Rothschild D, Wolfers J (2010) Forecasting elections: Voter inten-
tions versus expectations. Working paper, National Institute of
Economic Research, Cambridge, MA.
Satopää VA, Jensen ST, Mellers BA, Tetlock PE, Ungar LH (2014a)
Probability aggregation in time-series: Dynamic hierarchical
modeling of sparse expert beliefs. Ann. Appl. Statist. 8(2):
1256–1280.
Satopää VA, Baron J, Foster DP, Mellers BA, Tetlock PE, Ungar LH
(2014b) Combining multiple probability predictions using a
simple logit model. Internat. J. Forecasting 30(2):344–356.
Servan-Schreiber E (2007) About competitive forecasting. White
paper, NewsFutures, Paris.
Servan-Schreiber E (2012) Prediction markets: Trading uncertainty
for collective wisdom. Landemore H, Elster J, eds. Collective
Wisdom: Principles and Mechanisms (Cambridge University Press,
New York).
Servan-Schreiber E, Wolfers J, Pennock DM, Galebach B (2004)
Prediction markets: Does money matter? Electronic Markets
14(3):243–251.
Sherden WA (1998) The Fortune Sellers: The Big Business of Buying and
Selling Predictions (John Wiley & Sons, Hoboken, NJ).
Silver N (2012) The Signal and the Noise: Why So Many Predictions
Fail—But Some Don’t (Penguin, New York).
Snowberg E, Wolfers J (2010) Explaining the favorite-long shot bias: Is
it risk-love or misperceptions? J. Political Econom. 118(4):723–746.
Sonnemann U, Camerer C, Fox C, Langer T (2013) How psychological
framing affects economic market prices in the lab and field. Proc.
Natl. Acad. Sci. USA 110(29):11779–11784.
Spann M, Skiera B (2003) Internet-based virtual stock markets for
business forecasting. Management Sci. 49(10):1310–1326.
Surowiecki J (2005) The Wisdom of Crowds (Anchor Books, New York).
Swets J (1996) Signal Detection Theory and ROC Analysis in Psychology
and Diagnostics: Collected Papers (Lawrence Erlbaum Associates,
Mahwah, NJ).
Tetlock P (2005) Expert Political Judgment: How Good Is It? How Can
We Know? (Princeton University Press, Princeton, NJ).
Tetlock PE, Mellers BA, Rohrbaugh N, Chen E (2014) Forecasting
tournaments: Tools for increasing transparency and improving
the quality of debate. Current Directions Psych. Sci. 23(4):290–295.
Wolfers J (2009) Prediction markets: The collective knowledge of
market participants. CFA Institute Conf. Proc. Quart. 26(2):37–44.
Wolfers J, Zitzewitz E (2006) Interpreting prediction market prices as
probabilities. NBER Working Paper 12200, National Bureau of
Economic Research, Cambridge, MA.
Zou H, Hastie T (2005) Regularization and variable selection via the
elastic net. J. Royal Statist. Soc. 67(2):301–320.