How difficult is it to forecast the world’s most successful

products?
The case of the Apple iPhone
Petropoulos Fotios1,*, Nikolopoulos Konstantinos2, Litsa Akrivi3,
Polychronakis Yiannis4, Assimakopoulos Vassilios3
1
Lancaster Centre for Forecasting, Lancaster University Management School, Lancaster
University, Lancaster, UK
2
Bangor Business School, Bangor University, Bangor, UK
3
Forecasting and Strategy Unit, School of Electrical and Computer Engineering, National
Technical University of Athens, Greece
4
University of Salford, Manchester, UK

f.petropoulos@lancaster.ac.uk; k.nikolopoulos@bangor.ac.uk; alitsa@fsu.gr;

y.polychronakis@salford.ac.uk; vassim@fsu.gr

*corresponding author

Abstract
Apple is the largest publicly traded company in the world by market capitalization, and also
the largest technology company in the world by revenue and profit. Accurate forecasting unit
sales is of major importance as even the smallest of errors will have a huge impact in terms of
sales management and revenues. The current research investigates extrapolation alternatives
relating to iPhone unit sales, probably the corporation’s most influential product. Simple and
straightforward time series techniques (naïve, exponential smoothing methods, theta model)
are competing judgmental approaches (unaided judgment, interaction groups) performed by
three groups of participants in terms of expertise (novices, semi-experts, experts). Moreover,
two subgroups (for each level of expertise) were formed, regarding the amount of information
provided to participants. The performance of the alternatives was measured in terms of bias
and accuracy, using simple error and absolute percentage error. Results indicate that
judgmental approaches produced greater forecasts than quantitative methods, whilst focusing
effect is verified to lead to cognitive bias when forecasting future outcomes. As far as
performance is concerned, a combination of time series extrapolation techniques and
judgmental forecasts derived from groups with limited knowledge produces the smallest
forecasting errors.
Keywords: sales forecasting, Apple, iPhone, statistical forecasting methods, judgmental
forecasting

1. Introduction
Sales forecasting is an integral part of every aspect of the company, serving as an essential
tool for planning, budgeting, inventory and personnel planning (Makridakis & Wheelwright,
1989). Fildes & Hastings (1994) emphasize the importance of sales forecasting, providing the
short and long term effects of its absence. The significance of sales forecasting becomes even
greater when multinational corporations are considered. Apple is the largest publicly traded
company in the world by market capitalization, featuring worldwide annual revenues in 2010
of $65 billion, and growing to $108 billion in 2011 (Apple Form 10-K, 2010), more than
Google and Microsoft combined. Apple’s hardware product list includes the widely known
Mac (line of computers), the iPod (portable music device), the iPhone (smartphone) and the
iPad (tablet computer). On the other hand, its software consists mainly of two operating
systems (Mac OS X and iOS, oriented towards desktop and mobile systems respectively), as
well as some multimedia and productivity suites.

Apple’s trading share value raised from approximately $8 in early 2001 to $100 in mid 2007.
Since then, its growth was almost exponential and even surpassed the $600 limit (April,
2012). This growth is due to three primary reasons: the deal with Intel (2006) as the main
supplier of CPUs for the Mac series, the introduction of the iPhone and the iPad, and the
growth of the iOS mobile operating system along with the Apple Store service, which enabled
the distribution of third-party software applications. The iPhone alone, with more than 200
million units sold since its launch, is considered the major hit of the company, whilst being
the benchmark smartphone for other manufacturers. Forecasting sales of the iPhone is a very
challenging task, taking into account the fact that Apple is very strict about its methods of
tightly controlling information regarding product launches and updates.

The current study investigates and compares alternative statistical and judgmental forecasting
methods when considering the iPhone quarterly sales. In more detail, the performance of
simple and robust time series methods (such as damped exponential smoothing and the Theta
model) is compared to the accuracy of individual and group judgmental approaches. The
research aims to provide insights into the relationship between forecasting accuracy and level
of expertise, the effect of anchoring on judgmental approaches and the effectiveness of
combined forecasts.

2. Literature review and research questions

Statistical forecasting methods are considered suitable for extrapolating time series, assuming
that the established patterns and relationships recorded will not change during the required
forecasting horizon. Simple approaches, such as Single Exponential Smoothing (SES) and
Damped Exponential Smoothing (Gardner, 1985), have proven to be robust for a wide range
of data generating processes (Chatfield et al., 2001; Hyndman, 2001). Moreover, another
promising approach is the Theta model (Assimakopoulos & Nikolopoulos, 2000), a
framework that decomposes the original series into two or more “theta lines” and extrapolates
them separately. The Theta model achieved superior performance in the M3 forecasting
competition (Makridakis & Hibon, 2000), where data from various sources were considered.
On the other hand, judgmental approaches have proven to be crucial in sales forecasting. This
is mainly due to the fact that forecasters are able to integrate non time series information into
the forecasts (Webby & O’Connor, 1996), especially when key products are examined
(Edmundson et al., 1988). This will form the basis for the first research question:

Q1. Do judgmental approaches produce more accurate forecasts than statistical methods in
the case of the Apple iPhone?

One main disadvantage of judgmental forecasting approaches is the effect of anchoring,

which can be interpreted as cognitive bias, describing the human tendency of relying on one
piece of information when making judgments. Judgmental forecasting can be modeled as
anchor and adjustment (Tversky & Kahneman, 1974; Lawrence & O’Connor, 1992, 1995),
where a first approximation is produced (anchor) and then adjustments are made based on
additional information. The current study explores the impact of the amount of information
given to forecasters on forecasting performance:

Q2. How is the anchoring effect reflected in forecasting accuracy when additional
information is available?

Another important aspect of judgmental forecasting methods is the selection of

participants/forecasters with the appropriate level of experience. Experience is generally
divided into technical and causal knowledge. Technical knowledge is defined as knowledge
of statistical methods and the biases inherent in human judgment. When time series are
considered, Edmundson (1990) and Sanders & Ritzman (1992) concluded that technical
knowledge does not improve forecasting accuracy. On the other hand, Lawrence et al. (1985)
provided evidence that technical knowledge is beneficial when the data are presented in a
numerical format. Furthermore, the literature evidence of the value of expertise is conflicting,
suggesting in some cases that novices perform equal to or better than experts (for example
Önkal & Muradoglu, 1994; Thomson et al., 2004). Additionally, it is suggested that expertise
beyond a certain minimal level has limited value (Armstrong, 1985). As a result, there is no
clear evidence of the relation between experience and forecasting effectiveness:

Q3. What is the relation between the level of expertise and forecasting accuracy when
successful products are considered?
The literature has demonstrated strong empirical evidence of the superiority of the
combination of statistical and/or judgmental approaches in relation to forecasting
performance (Makridakis et al., 1982; Makridakis & Winkler, 1983; Clemen, 1989;
Surowiecki, 2005). In addition, averaging reduces the variance in forecasting errors and
therefore the uncertainty in predictions, rendering the selection of combinations less risky
than individual methods (Hibon & Evgeniou, 2005). Whilst many studies have focused on
sophisticated weight selection processes (for example Jose & Winkler, 2008; Kolassa, 2011),
simple averaging is considered effective in most cases, especially in the presence of
uncertainty. The last research question refers to the performance of combined forecasts in the
context of major product sales:

Q4. Is the combination of statistical and judgmental methods beneficial in forecasting the
sales of successful products?

3. Methodology
3.1 Empirical data
Fiscal quarterly sales (in millions) for Apple iPhone (Apple Earnings Releases, 2012) are
presented graphically in Figure 1, along with its linear trend line. The most dominant pattern
in the series is an almost linear trend, except for the very last observation. The unusually high
demand for fiscal quarter 2012 Q1 is most probably due to the latest update of the product
(iPhone 4S) which occurred in October 2011. Similar updates were made earlier (June 2009
for iPhone 3GS and June 2010 for iPhone 4) but the recorded demand in the corresponding
fiscal quarters was not as unusual. Moreover, the correlograms for ACF and PACF are
presented in Figure 2. It is obvious that the seasonality component is absent from this specific
series, so iPhone can be described as a non-seasonal product. The required forecasting
horizon was set to four quarters, equal to an entire year. Unfortunately, the results of this
paper are limited to the next fiscal quarter, as no other financial results have been released by
the company.

3.2 Statistical methods

Exponential smoothing techniques are widely used for managerial purposes and sales
forecasting. Three widely used exponential smoothing models are tested in this study. SES is
the simplest exponential smoothing approach, where no trend component is assumed. Holt
exponential smoothing (Holt, 1957), expands SES by adding a trend component. Holt’s
method was proven to be biased in the long run, so a dampening factor for the trend
component was proposed (Gardner, 1985), resulting in the Damped method. Lastly, we are
considering the very promising Theta model, which is almost equivalent to SES with drift
(Hyndman & Billah, 2003) in the case that its classic approach is considered (Nikolopoulos et
al., 2012). The classic Theta model refers to the decomposition of the initial series into two
symmetrical theta lines with parameters equal to zero and two, which are extrapolated with
linear regression and SES respectively. The final ‘theta’ forecast is the simple combination of
the two extrapolated lines. The forecasts of the statistical methods used in this research were
produced using the free statistical software R (R Development Core Team, 2012) and
specifically the forecast package (Hyndman, 2012).

Figure 1. Apple iPhone unit sales (in millions) per fiscal quarter

Figure 2. ACF and PACF plots of the data

3.3 Judgmental methods

Two judgmental approaches were considered for the current research. Unaided Judgment (UJ)
is the simplest and most popular forecasting approach. Forecasts are submitted by the
participants without the use of any structured method or guidance. Individual forecasts
gathered from UJ implementations are usually averaged so as to produce a group forecast.
The Interaction Groups approach (IG) is widely used in practice (Lawrence et al., 2000) and
suggests the active interaction of a group of experts until a consensus forecast is reached
through debate and discussion. In this way, all members of the group have access to a wider
range of information. Literature suggests that group methods are generally more accurate than
averaged individual forecasts (for example Ang & O’Connor, 1991), even if group discussion
may lead to optimism (Brenner et al., 2005).

The current research investigates the performance of judgmental forecasting accuracy for
three levels of expertise. The lowest level of expertise (novices) consisted of 39
undergraduate students from the Electrical and Computer Engineering School of the National
Technical University of Athens (NTUA) and the experiment took place during the very first
lecture of the “Forecasting Techniques” course. Participants of the second level (semi-
experts) were 15 postgraduate students from the NTUA and the procedure was carried out
during the “Decision Support Systems” course of the “Environment & Development”
graduate program. Lastly, the group with the highest level of expertise was completed by 11
members and postgraduate students from the “Decision Support Systems Laboratory” and the
“Forecasting & Strategy Unit” at NTUA. As a result, in the current study expertise is
primarily identified as knowledge of the technical aspects of time series analysis and
forecasting. Participants from each of the three groups were divided into two smaller groups,
so that a different amount of information was available to each subgroup. Each individual
would know exactly one of the following about the nature of the available time series:
“Apple iPhone unit sales” or “Smartphone unit sales”
Table 1 demonstrates the group names and the number of participants in each group in
relation to their level of expertise as well as the nature of the information available. The same
data from Apple iPhone quarterly unit sales were presented to all participants graphically and
a simple questionnaire was completed relating to their individual unaided judgmental
forecasts for each one of the upcoming four fiscal quarters.

Table 1. Group names and number of participants per group

“Apple iPhone unit sales” “Smartphone unit sales”
Level of
Number of Number of
Expertise Group name Group name
participants participants
Novices N1 18 N2 21
Semi-Experts S1 8 S2 7
Experts E1 6 E2 5

Upon the completion of UJ, the participants of the groups of semi-experts and experts (S1,
S2, E1 and E2) were gathered together towards the formulation of the IG. Access to
previously submitted forecasts was granted to all participants in each group, whilst they were
asked to debate, discuss and argue their viewpoints. Moreover, individuals with
minimum/maximum personal judgments were requested to provide justification to the other
panelists. The meetings were generally short (30 to 40 minutes) and ended whenever
participants came to a consensus on the forecasts for the upcoming four fiscal quarters.

3.4 Measuring performance

The evaluation of statistical and judgmental approaches was performed by measuring the bias
and the forecasting accuracy of the alternatives. Simple error (E) is the simplest bias metric,
denoting the signed difference of actual recorded values to the single produced forecast or the
averaged individual forecasts of the judgmental approaches:

Error for single forecast: E =Y −F

1 n
Error for averaged individual forecasts: E =Y − ∑ Fi
n i =1
When accuracy is considered, any of the widely used accuracy metrics is suitable, as only one
series is examined in this study. The absolute percentage error (APE) fits the needs of this
exercise, providing, at the same time, easy-to-interpret results:

Y −F F
Absolute percentage error for single forecast: APE = = 1 − (%)
Y Y
1 n n
Y− ∑ Fi ∑ Fi
Absolute percentage error for averaged n i =1 i =1
APE = = 1− (%)
individual forecasts: Y n ⋅Y

4. Results and discussion

The produced forecasts and respective values of accuracy metrics, in terms of signed error
and absolute percentage error, for all alternative approaches considered in the current study
are presented in Table 2. As mentioned earlier, the presented results are limited to the second
fiscal quarter of 2012, due to the fact that no other economic results are yet available. In fact,
the recorded sales of Apple iPhone for Q2 2012 are 35064 thousand units. The results are
divided into three main categories. In the first category, results for the statistical methods
(SES, Holt, Damped and Theta) are presented. The second category includes the two
judgmental approaches (Unaided Judgment and Interaction Groups) for each level of
expertise (novices, semi-experts and experts) and each group (with regards to the volume of
information provided to participants). The last category refers to the combinations tested in
this exercise, where the best blend between each judgmental group and statistical methods is
presented. Forecasts and signed error values refer to thousands of product units. Lastly,
forecasts and simple errors are rounded to the first decimal point, whilst absolute percentage
errors are rounded to the second decimal point.

Table 2. Statistical accuracy for different approaches

Expertise Level Method Group Forecast E APE

SES 32980.8 2083.2 5.94%
Statistical
Methods

Holt 31276.6 3787.4 10.80%

Damped 31185.9 3878.1 11.06%
Theta 34291.8 772.3 2.20%
N1 39875.0 -4811.0 13.72%
Novices UJ
N2 38434.8 -3370.8 9.61%
S1 37375.0 -2311.0 6.59%
UJ
S2 38833.3 -3769.3 10.75%
Approaches
Judgmental

Semi-Experts
S1 40000.0 -4936.0 14.08%
IG
S2 40000.0 -4936.0 14.08%
E1 40416.7 -5352.7 15.27%
UJ
E2 37750.0 -2686.0 7.66%
Experts
E1 42000.0 -6936.0 19.78%
IG
E2 37500.0 -2436.0 6.95%
Damped & UJ N1 35530.5 -466.5 1.33%
Novices
Holt & UJ N2 34855.7 208.3 0.59%
SES & UJ S1 35177.9 -113.9 0.32%
Best Combinations

Holt & UJ S2 35055.0 9.0 0.03%

Semi-Experts
Damped & IG S1 35593.0 -529.0 1.51%
Damped & IG S2 35593.0 -529.0 1.51%
Damped & UJ E1 35801.3 -737.3 2.10%
SES & UJ E2 35365.4 -301.4 0.86%
Experts
Damped & IG E1 36593.0 -1529.0 4.36%
SES & IG E2 35240.4 -176.4 0.50%

A close study of the results presented in Table 2 leads us to some remarkable observations. In
the case in which individual techniques are examined, it is noted that all statistical methods
result in positive simple errors, which denote that sales forecasts are lower in value than the
actual observed sales for Q2-2012 fiscal quarter. On the other hand the same metric has a
negative sign for all judgmental approaches, meaning that participants of all groups were, on
average, optimistic regarding the forecasted sales. Moreover, if APE values are examined, it
is easily interpreted that Theta had the best performance amongst all other individual
techniques (2.20%), whilst the second best performance is recorded for the SES method
(5.94%). On average, statistical approaches were less biased (2630.2 to -4154.5) and more
accurate (7.50% to 11.85%) than judgmental ones.

The next observation comes from the comparison of the accuracy of different groups when
the amount of available information is examined. As previously mentioned, groups N1, S1
and E1 were exposed to the full complement of available data (unit sales of Apple iPhone),
whereas groups N2, S2 and E2 were just told that “smartphone” unit sales were being
explored. It is clear that for all novice and expert groups the extra information led to less
accurate and more biased forecasts, whilst the picture was reversed for the semi-experts.
When all groups are examined, it is clear that the extra information resulted in a deterioration
of the participants’ forecast performance for both E and APE (-4869.3 to
-3439.6 and 13.89% to 9.81% respectively).

Figure 3. Box-Whiskers display for unaided individual forecasts of all judgmental

approaches per group

Another interesting result lies in the performance comparison between the various levels of
expertise and the different judgmental approaches. Even if accuracy improves slightly when
“semi-experts” are preferred to “novices” (11.37% to 11.67%), there is a significant (on
average) deterioration in the performance of “experts” (12.41%), verifying previously
published results (Armstrong, 1985). This result is mostly due to the performance of group
E1, which had the worst performance amongst all other groups for both approaches (UJ and
IG). Regarding the performance of the group forecasting approach (IG) versus the individual
unaided judgments, the results indicate that IG performs worse in 3 out of 4 cases. The worst
performance is recorded both in terms of bias, where the IG approach resulted in more
optimistic forecasts than both the average and median of individuals, and accuracy, as a result
of the APE rising from 10.07% for UJ to 13.72% for IG. The only exception is observed in
group E2, which is the one group that takes advantage of the grouping approach. In an
attempt to interpret this result, a Box-Whiskers plot for the individual unaided forecasts of all
groups is illustrated in Figure 3. It is clear that forecasts of individuals in group E2 had the
largest dispersion amongst all other groups, and especially in comparison with groups S1, S2
and E1. Therefore a discussion procedure (IG approach) between the participants of group E2
was beneficial, resulting in more accurate forecasts.

Lastly, the performance of various combinations between statistical and judgmental

approaches is examined. The pessimistic forecasts produced by all statistical methods doubled
with the optimistic sales forecasts from judgmental approaches intuitively denote improved
performance when these two alternatives are combined. Table 2 demonstrates the best
combination of statistical and judgmental approaches for each one of the groups considered.
In fact, simple combinations with equal weights result in almost every case (except for E1 -
Damped & IG group) in more accurate and less biased forecasts compared to single statistical
or judgmental approaches. The recorded MAPE for the best combinations is on average just
1.31%, whilst 5 out of 10 combinations are more accurate than 1%. The best performance is
observed for the Holt & UJ approach in group S2 where the APE is as low as 0.03%. In more
detail, the best performance is achieved by a combination of a statistical method suitable for
trended time series and the mean unaided judgments of participants with medium level of
experience where no additional information, other than “smartphone unit sales”, was
available.

5. Conclusions and future research

The current study examined the comparative forecasting performance of alternative statistical
and judgmental approaches when the quarterly sales for Apple iPhone were considered. The
conclusions of the study can be summarized as follows:
• When single approaches are explored, statistical extrapolations are less biased and
more accurate than judgmental ones. Moreover, judgmental forecasts suffer from
serious optimistic biases.
• Additional information regarding the full nature of the data is not always beneficial
and may act as anchoring for participants. This result may be associated with the
profile of the company considered in this study.
 • Simple and individual judgmental approaches (UJ) should be preferred to more
complicated ones (IG), unless individual positions suffer from great discrepancies.
• A minimum increase in the expertise level (for example, postgraduate students as
opposed to undergraduate ones) seems to be beneficial.
• Combinations of statistical and judgmental approaches have once more achieved
great accuracy levels, surpassing, in most cases, the performance of individual
methods.
In any case, the study should not be considered as complete unless all four fiscal quarters’
actual values become available. A possible avenue for future work should be the replication
of this study with the rest of Apple’s products (iPod, iPad and Mac), in which additional time
series characteristics (such as seasonality) are observed. Moreover, alternative time series
data, for example smartphone unit sales of other manufacturers, should be examined. Lastly,
the performance of structured judgmental approaches (such as Structured Analogies) should
also be investigated.

References
Ang S and O’Connor M (1991). The effect of group interaction strategies on performance in
time series extrapolation. International Journal of Forecasting 7(2): 141-149.
Apple Form 10-K (2010). URL: http://investor.apple.com/secfiling.cfm?filingID=1193125-
11-282113&CIK=320193 [Accessed May 1, 2012].
Apple Earnings Releases (2012). URL: http://investor.apple.com/results.cfm [Accessed May
1, 2012].
Armstrong JS (1985). Long-Range Forecasting: From Crystal Ball to Computer (2nd ed).
Wiley: New York.
Assimakopoulos V and Nikolopoulos K (2000). The theta model: A decomposition approach
to forecasting. International Journal of Forecasting 16(4): 521-530.
Brenner L, Griffin D and Koehler DJ (2005). Collaborative planning and predication: Does
group discussion affect optimistic case-based judgement. Organizational Behavior and
Human Decision Processes 97(1): 64-81.
Chatfield C, Koehler AB, Ord JK and Snyder RD (2001). A new look at models for
exponential smoothing. The Statistician 50(2): 147-159.
Clemen RT (1989). Combining forecasts: A review and annotated bibliography (with
discussion). International Journal of Forecasting 5(4): 559-583.
Edmundson R, Lawrence M and O’Connor M (1988). The use of non time series information
in sales forecasting: A case study. Journal of Forecasting 7(3): 201-211.
Edmundson R (1990). Decomposition: A strategy for judgmental forecasting. Journal of
Forecasting 9(4): 305-314.
Fildes R and Hastings R (1994). The organization and improvement of market forecasting.
Journal of the Operational Research Society 45(1): 1-16.
Gardner Jr ES (1985). Exponential smoothing: The state of the art. Journal of Forecasting
4(1): 1-38.
Hibon M and Evgeniou T (2005). To combine or not to combine: selecting among forecasts
and their combinations. International Journal of Forecasting 21(1): 15-24.
Holt CC (1957). Forecasting seasonals and trends by exponentially weighted averages. ONR
Memorandum 52/1957. Pittsburgh: Carnegie Institute of Technology. Reprinted with
discussion in 2004. International Journal of Forecasting 20(1): 5-10.
Hyndman RJ (2001). It is time to move from what to why. International Journal of
Forecasting 17(4): 567-570.
Hyndman RJ and Billah B (2003). Unmasking the Theta method. International Journal of
Forecasting 19(3): 287-290.
Hyndman RJ (2012). Forecast: forecasting functions for time series. R package version 3.22.
URL: http://CRAN.Rproject.org/package=forecast [Accessed June 10, 2012].
Jose VRR and Winkler RL (2008). Simple robust averages of forecasts: Some empirical
results. International Journal of Forecasting 24(1): 163-169.
Kolassa S (2011). Combining exponential smoothing forecasts using Akaike weights.
International Journal of Forecasting 27(2): 238-251.
Lawrence M, Edmundson R and O’Connor M (1985). An examination of the accuracy of
judgmental extrapolation of time series. International Journal of Forecasting 1(1): 25-35.
Lawrence M and O’Connor M (1992). Exploring judgmental forecasting. International
Journal of Forecasting 8(1): 15-26.
Lawrence M and O’Connor M (1995). The anchoring and adjustment heuristic in time series
forecasting. Journal of Forecasting 14(5): 443-451.
Lawrence M, O’Connor M and Edmundson B (2000). A field study of sales forecasting
accuracy and processes. European Journal of Operational Research 122(1): 151-160.
Makridakis S, Andersen A, Carbone R, Fildes R, Hibon M, Lewandowski R, Newton J,
Parzen E and Winkler R (1982). The accuracy of extrapolation (time series) methods: results
of a forecasting competition. Journal of Forecasting 1(2): 111-153.
Makridakis S and Winkler R (1983). Average of Forecasts: Some Empirical Results.
Management Science 29(9): 987-996.
Makridakis S and Wheelwright SC (1989). Forecasting Methods for Management (5th ed).
Wiley: Chichester.
Makridakis S and Hibon M (2000). The M3-Competition: results, conclusions and
implications. International Journal of Forecasting 16(4): 451-476.
Nikolopoulos K, Assimakopoulos V, Bougioukos N, Litsa A and Petropoulos F (2012). The
Theta model: An essential forecasting tool for Supply Chain planning. In: Lee G (ed).
Advances in Automation and Robotics Vol. 2, Lecture Notes in Electrical Engineering 123:
431-437.
Önkal D and Muradoglu G (1994 ). Evaluating probabilistic forecasts of stock prices in a
developing stock market. European Journal of Operational Research 74(2): 350-358.
R Development Core Team (2012). R: a language and environment for statistical computing,
version 2.15.0. Vienna, Austria: R Foundation for Statistical Computing. URL: http://www.R-
project.org [Accessed June 10, 2012].
Sanders N and Ritzman L (1992). The need for contextual and technical knowledge in
judgmental forecasting. Journal of Behavioral Decision Making 5(1): 39-52.
Surowiecki J (2005). The Wisdom of Crowds. Anchor Book: New York.
Thomson ME, Pollock AC, Henriksen KB and Macaulay A (2004). The influence of the
forecast horizon on judgmental probability forecasts of exchange rate movements. The
European Journal of Finance 10(4): 290-307.
Tversky A and Kahneman D (1974). Judgment under uncertainty: Heuristics and
biases. Science 185(4157): 1124-1131.
Webby R and O’Connor M (1996). Judgmental and statistical time series forecasting: A
review of the literature. International Journal of Forecasting 12(1): 91-118.