ABN AMRO CLIENTS PREDICTING THE AEX-INDEX VALUE

Two comparisons: how accurate are the expectations compared to

the random walk model and the perspective on decisions of others?

Emmely Wildeboer, 296414

August, 2010

In this study we investigate to what extent the predictions made by active banking clients of ABN
Amro on the AEX value are accurate. The predicted values by the respondents of the questionnaire
are compared to two different scenarios: (i) approach according to the ‘traditional’ random walk
model and (ii) what they think others will predict. Each comparison is made using three methods: the
mean-squared error, the mean absolute error and the mean percentage error. In the first
comparison, the random walk approach finds more support. This is not to say that the AEX follows a
random, it suggests in our case that the random walk model makes more accurate predictions than
the respondents do. The results of second comparison recommends that respondents are better off
when they predict what they think others will predict, instead of making predictions on their own.

Thesis International Bachelor in Economics and Business Economics

Erasmus School of Economics

Behavioral Finance

Supervisor: Dr. Remco Zwinkels

 The most predictable thing about the stock market is the number of experts who take credit for
predicting it.” -Dave Weinbaum

1. Introduction

Making predictions on the stock markets is one of the greatest challenges an investors faces. In order
to acquire high returns from stock investment, the timing of buying and selling assets is essential.
The principle of ‘equity market timing’ in finance refers to the fact that one should issue shares at
high prices and repurchase the shares when their prices are low.
In addition to market timing, predicting market values is important since it is a good indicator for the
investment choice. In order to make profits, it makes perfect sense that the expectations should
come close to the actual values. The stereotype investor on the stock market is not considered to be
clairvoyant, and makes therefore use of different tools to come closest to the actual value. The stock
market involves enormous returns, which makes predictions crucial for successful investment.
Naturally individuals are delighted when predicting correctly since it will lead to positive returns,
though this is not always the case. In making predictions, McKenzie and Amin (2002) support several
hypotheses, of which the most peculiar and important explains that incorrect predictions (whether
they are from rare or from common events) are more supported than correct predictions. Making
predictions on rare events are called bold predictions, and McKenzie and Amin show the importance
of this boldness in the support shown to the correctness. Nevertheless, in the stock market correct
predictions are preferred, whether they are supported or not. Why is predicting in the stock market
so special? Stock time series have various characteristics that contribute to the fact that the
prediction task is rather uncommon and calls for special concern. Hellström and Holmström (1998)
explain the specific properties;

 Stock prediction is generally believed to be a very difficult exercise. The task of predicting
stocks can be compared to that of inventing a perpetuum mobile1.
 The process behaves a lot like a random-walk process, in the sense that the autocorrelation
for day to changes is rather low.
 The fundamental process is time varying, meaning the process is ‘regime shifting’. As the
stock markets move from one period to another, the level of noise and volatility change.
Hence, this causes problems for traditional algorithms for time series predictions.

1
A perpetuum mobile is a devise or motion that, once is has started, were to continue indefinitely.

2
 1.1 Tools in predicting on stock market
In predicting stock market prices or values, tools are used to forecast expected values and to make
trading decisions. In 1970, Fama published an article that would become of major importance for
financial economists. In his “Efficient Market Hypothesis”, Fama introduced a model that states that
stock markets fully reflect all information available. Stock markets assumed to follow a random walk
like mentioned above, in the sense that today’s stock market prices are independent of those from
yesterday. In 1990, Delong, Shleifer, Summers and Waldmann argued that investors are exposed to
sentiment. Behavioral economics soon began to show its link with making predictions in the financial
world. A new generation of techniques and tools became known to intelligently support people in
analyzing data, finding valuable knowledge and in some cases performing analysis automatically.
These techniques and tools include approaches like fuzzy times series (using linguistic variables to
deal with the vagueness of human knowledge2) and Naïve Bayes. The Naïve Bayes uses a classifier
assuming that the presence or absence of a particular property is not related to the presence or
absence of any other in future time.
In spite of the traditionally enormous support for the Efficient Market Hypothesis, most market
players now believe they can, at least partially, predict market values. Different sentiment indices
incorporate investor sentiment into a tool for investment decisions. For example, the Nova-Ursa ratio
is an indicator that uses the Rydex Nova and Rydex Ursa mutual funds to include sentiment. On the
other hand, the Put/Call ratio gauges sentiment by dividing the number of traded put options by the
number of traded call options. Market players buy whenever the market is ’bullish’ and sell
whenever the market is ‘bearish’; the basics of a sentiment index.

1.2 The ABN Amro sentiment index

In making predictions on the stock market, investors make use of many sources that can be helpful in
predicting correctly. In addition to financial gurus and experts, a frequently used tool is a so called
sentiment index. The first sentiment index was created in 1963 by the New York based Investor
Intelligence. This index shows the ratio of the number of investment advisors who are bearish to the
total number of advisors who are either bearish or bullish. In this context, we use bearish to name
investors who drive down prices relative to the fundamental value. On the other hand, bullish
investors are more optimistic; they drive up their prices. The outcome of the index tells the investor
that one should buy when investment advisors are bearish and sell when advisors are bullish. A
sentiment index is promoted as a contrary indicator; when the index turns bearish, the investors turn
bullish.

2
Dubois and Prade, 1990

3
The data provided by ABN Amro in this study is obtained through survey data and is incorporated
into a trading index; what percentage of the respondents think the AEX value will be higher or lower
than the last communicated value?
Based on the data, this study investigates to what extent the respondents were correct in predicting
the AEX value. To test the accuracy of the predictions made by the respondents, two comparisons
are made which include two approaches derived from: the random walk model and the phenomenon
herd behavior in behavioral finance. To be able to make the comparisons between the different
scenarios, various measures are used: the mean squared error, the mean absolute error and the
mean percentage error. For these methods the lowest values are preferred.
Firstly, the outcomes of the respondents (scenario 1) are compared to the outcomes if the AEX value
follows a random walk (scenario 2), a heavily supported theory in the field of predicting stock prices.
We apply the random walk model to our study by using the actual value of the date prior to the
forecast date, therefore assuming no change in the AEX value. The results indicate that the random
walk approach finds more support, since the values for MSE, MAE and MPE are lower in this can than
in the case with the predicted values by the respondents.
Secondly, we compare the prediction made by the respondents to the values they think that other
respondents will predict (scenario 3). Herd behavior exist in financial decision-making because
individuals think others can make better decisions than themselves. In the second comparison, we
can observe whether the accuracy differs in these two cases; do the respondents make more
accurate predictions for themselves or for others? The outcome of the second comparison is
surprising since the respondents make more accurate predictions for others than on behalf of
themselves.
The outline of this paper is as follows: in section 3 the data is described and given, followed by the
methodology used in section 4. The results of the study is given in section 5. Ultimately, section 6
contains the conclusion and discussion part.

2. Relevant literature

Many theories and models have been created to investigate the changes in the stock market. These
theories predict how stock prices change and what influences these changes.

2.1 Efficient Market Hypothesis and the Random Walk

Stock prices have long been assumed to reflect all information available, as is explained by the
Efficient Market Hypothesis. In an efficient market, security prices fully reflect all obtained

4
information. When new information becomes available, prices are adjusted without any delay.
Therefore, no arbitrage opportunities exist that would give investors the opportunity to gain above-
average returns without accepting above-average risk.
The Efficient Market Hypothesis gives explanation to the random walk model by hypothesizing the
price changes arise as a result of changes in information regarding the security in question.
The idea of the random walk model is that the flow of information is unrestrained and that
information is instantaneously reflected in stock prices. The price change of tomorrow only reflects
tomorrow’s news and is independent of today, so no change occurs in the price. Hence, the future
path of the price level of a security is no more predictable than the path of a series of cumulated
random numbers (Fama, 1965). In 1973, Malkiel explained the random walk theory in an interesting
and unusual way. He used blindfolded chimpanzees to throw darts at the Wall Street Journal and
concluded that the animals could select a portfolio that would do just as well as the financial experts.

Nowadays, it is often believed that the stock market is at least to some extent predictable. Since the
1990s, the intellectual dominance of the EMH became less common. In the early 2000’s two
arguments against the EMH arose. Firstly, periods of market irrationality occurred. It became
recognized that if one could avoid the psychological drawbacks that investors are prone to, then it
must be possible to outperform the market. Secondly, the tendency of markets to overreact makes
the stock markets at least somewhat predictable. The earliest observations of symptoms of
overreaction came from J.M. Keynes, who argued that “….day-to-day fluctuations in the profits of
existing investments, which are obviously of an ephemeral and no significant character, tend to have
an altogether excessive, and even absurd, influence on the market” (Keynes, 1936).
In a paper by Kahneman and Tversky in 1982, it is explained that people are inclined to overweight
recent information, while underweighting prior data. Individuals base their decision-making on a
rule-of-thumb: “The predicted value is selected so that the standing of the case in the distribution of
outcomes matches its standing in the distribution of impressions.” This implies that in making
predictions, the extremeness of predictions must be weakened by considerations of predictability.
The relevance of the behavior of overreaction in economic forecasting on the stock market is
underlined in for example De Bondt and Thaler (1985) and Lehmann (1990).

2.2 The influence of behavioral economics on financial decision making

Like mentioned before, periods of market irrationality and overconfidence weakened the support for
the Efficient Market Hypothesis. The EMH had increasing difficulty in explaining periods during for
example the Dot-Com Bubble. During the Dot-Com Bubble a synthesis of increasing stock prices,

5
market confidence, individual speculation in stocks (speculation in the sense that it did not promise
any safety), and widely accessible venture capital gave rise to an environment in which investors
disregarded traditional metrics like the price-to-earnings ratio. The development of this bubble led to
many bankruptcies and losses for investors.

In the financial literature, a new paradigm emerged based on behavioral economics. The alternative
model is based on two assumptions;

(i) Investors are exposed to sentiment, as mentioned first by DeLong, Shleifer, Summers and
Waldmann in 1990. Not all demand changes are based on rational behavior; some
responses to changes in expectations are not fully justified by information. An example
of such pseudo-signals that influence investors’ decisions, is advice of financial gurus and
forecasters. Investors that are influenced by such sentiment, are called ‘noise traders’.
The essence of noise in liquid markets was already stressed out by Fischer Black in 1986,
who argues that “noise trading is essential to the existence of liquid markets”. Noise in
this sense is information that has not appeared yet, it is the uncertainty about future
demand and supply conditions within and across sectors3.

(ii) The second assumption is laid out by Shleifer and Vishny in 1997 and it argues that
betting against sentimental investors is expensive and not without a risk. Therefore,
there are limits to arbitrage.

After discovering the influence of sentiment on investors, it is essential to pay attention to the effects
of sentiment on stocks. Since 1980, many research has been done on the influence of investor
sentiment. In these early studies, the influence of sentiment was left implicit, mainly because it was
hard to differentiate a random walk from a long-lived bubble. In more current studies, the tests show
stronger evidence on the influence of sentiment on investment. Two types of investors are
distinguished by DeLong, Shleifer, Summers and Waldmann (1990). Firstly, there are rational
arbitrageurs who are not influenced by sentiment. These traders are limited in the sense that time
horizons are short and trading and short selling is costly and risky. Secondly, irrational traders are
those who are in fact influenced by sentiment. Mispricing can arise when there is a combination of
events; changes in sentiment on the part of the irrational traders, and a limit to arbitrage from the
rational investors (because of short time horizons or from costs and risks of trading and short selling).

3
Black, F., “Noise”, 1986.

6
The influence of sentiment on stock return is positively correlated; when sentiment increases,
‘speculative’ (those with higher risks involved) stocks have higher returns.

2.3 The use of a sentiment index

Despite the fact that a sentiment index is often created to advice investors, research done by Solt
and Statman (1988) showed that a sentiment index is not useful as an indicator of future stock price
changes. The outcome is based on the research of data of 1000 observations from the year 1963, the
introduction of the used sentiment index, until 1985. The question that naturally arises afterwards, is
why sentiment is still regarded to be of major importance in financial decision making. Individuals
believe in the usefulness of a sentiment index because those individuals are not able to recognize
patterns in random data and ignore evidence that proves their counter beliefs. A striking example in
the failure to recognize randomness in data is the belief in the “hot hand” in basketball. This view is
taken by coaches, players and fans and implies that one has a “hot hand” when one experiences a
series of hits. People tend to believe that after scoring a hit, a player is more likely to score another
hit than a miss.
Another common belief says that when people commit errors, experience leads them to recognize
their faults and avoid them in the future. An appealing aspect of the belief in the “hot hand” is that
its strength increases with experience.
When we link the “hot hand” theory the respondents predicting AEX values, it can be suggested that
respondents who were correct in t=1, have a greater confidence in being correct in t=2. It is obvious
that its strength increases with experience; assume you have (nearly) correct expected values for 5
times in a row, then you will believe you will be correct the sixth time without a doubt.

The importance of investor sentiment on the stock market has been made very clear in numerous
studies. In addition, it can be questioned how to measure such sentiment? There many possible
methods for measuring sentiment, as laid out by Baker and Wurgler (2007). Several methods are:

- Investor surveys; to gain insight into the investors, one can ask how optimistic or pessimistic
investors are.
- Trading volume; liquidity can be viewed as an investor sentiment index. Trading volumes can
reveal underlying differences in opinion.
- IPO (initial public offering) First-Day returns; at times IPO’s earn exceptional returns on the
first day of trading, that it is hard to ignore the influence of investor enthusiasm.

7
 - Insider trading; corporate executives are better informed than outsiders concerning the
performance and true value of their firms. Therefore, the portfolios of those executives
unveil their opinions about mispricing of the related firm.

The data used in this paper, is obtained through the first method, namely investor surveys. ABN
Amro asked their banking clients to fill in a questionnaire on, among other things, predicting the AEX
value. The data derived from the questionnaire is used in this study.

3 Data

In this section the data used in this study is described. The data from ABN Amro is obtained through
surveys. As discussed in the previous section, investor surveys are valid methods to measure
sentiment amongst investors. For example, Pearce and Roley (1985), Brown and Cliff (2004) and Qiu
and Welch (2006) all used surveys to measure sentiment in a particular market.

3.1 Data description

In this study, the data is obtained from ABN Amro. Every two weeks, for a total of twelve times, ABN
Amro asked their active banking clients to fill in a questionnaire. Various topics were addressed in
the questionnaire, yet the following questions were asked specifically to be of interest for this study:

 What will be the AEX value at the time of our next survey number?
 What will be the minimum and maximum value for the AEX value on our next survey
number date?
 What do you expect that other ABN Amro clients will predict?

Subsequently, respondents were asked to fill in questions concerning other topics. In the appendix,
the questionnaire of survey number 5 is included, which asked additional questions on the oil
industry. However, due to irrelevance these additional questions are not taken into account.
Based on the obtained data, ABN Amro creates a trading index to measure the investor sentiment
amongst the ABN Amro investors. The trading index contains a grade, by means of how many
percent of the respondents thinks the future AEX value will be higher or lower than the current
value. An overview of the average forecasted AEX value per survey and the actual AEX value can be
given graphically:

8
 Comparison of predicted and actual AEX value
370
360
350
340
330 Average forecasted AEX value
320
Actual AEX value
310
300
290
1 2 3 4 5 6 7 8 9 10 11 12

Figure 1: Bar chart with the x-axis showing the survey numbers and the y-axis showing the value of the AEX.

As can be seen in figure one, most of the survey number the forecasted value differed not greatly
from the actual AEX value. However, in period 4, 6 and 10 the actual and forecasted value differed
enormously. The 12 surveys cover a period of from 29th of December 2009 until the 25th of May 2010.
ABN Amro asked their active banking clients4 to fill in the questionnaire on voluntary base.
Therefore, the number of respondents differ per survey number and the respondents are
anonymous. The sum of the respondents in the complete dataset is 1639.
Before the data can be used to make calculations to compare the prediction outcomes and the
random walk approach, a selection is made based on the following three criteria:

i. The respondent filled in the entire questionnaire.

ii. The maximum value given for the AEX value is larger than the minimum value.
iii. The predicted AEX value lies within the range given (minimum value < forecast < maximum
value).

In the provided dataset, the close AEX index at forecast date was missing for survey number 12 on
June the 4th 2010. The corresponded value is included and is naturally based on the actual value of
the AEX Index on this date, which can be found on www.iex.nl. The actual AEX value on June the 4th
of 2010 was 321,22.
After removing the 42 variables that did not follow the criteria, we end up with 1597 respondents.
Hence, an overview can be made of the data that will be used for calculations:

4
Active clients are considered those who have an amount of €100,000-€1,000,000 at their disposal

9
 Number of survey respondents ABN Amro questionnaire
250
200
150
100 Number of survey respondents
ABN Amro questionnaire
50
0
1 2 3 4 5 6 7 8 9 10 11 12

Figure 2: Bar chart with the x-axis showing the survey numbers and the y-axis showing the number of respondents per
survey number.

Based on figure 1, we can calculate that the average number of respondents is 133. When looking at
the shape of the bar chart, we see that total number of respondents per survey is declining. The
decline in respondents can for example be due to a decreasing interest in filling in the questionnaire,
but at this point there is too little information available on the respondents to make any conclusions
on this part.

4 Methodology

In this chapter, the methods used in this study are explained. In order to measure to what extent the
predicted values differ from the actual values, three formulas are applied to two scenarios.
Firstly the Mean Square Error (MSE), followed by the Mean Absolute Error (MAE) and finished by the
Mean Percentage Error (MPE).The mean squared error is a commonly used tool to compare different
outcomes. In 1981, Ohtani used a MSE comparison to test whether proxy variables are good
variables to make predictions. Wolski (1998) measures the accuracy of different forecasting methods
by reporting both the mean square error and the mean percentage error. Based on these previous
studies we assume that using mean errors is a valid method to compare the outcomes of our
respondents and the random walk model.
The mean error methods are firstly used to compare the scenarios: (i) taking the predicted AEX
values from the respondents, and (ii) taking the actual AEX values on the prior date, which are the
values communicated to the respondents. When the AEX values follow a random walk, it is not
possible to predict future values by taking into account past patterns. Therefore, we use values that
were communicated to the respondents because we assume a zero change in the AEX value.
Secondly, the mean error methods are applied to compare (i) the predicted AEX values from the

10
respondents with (iii) the predicted values that reflect the respondents’ opinion on what others will
predict.
The three mean error methods are based on the error, which is explained as:

Error ei = fi − yi, where fi is the prediction and yi the true value

To investigate to what extent the respondents of the survey were correct in predicting the AEX value,
the following three methods are used to compare the outcomes:

A. Mean Square Error

MSE = ∑ [(forecasted-actual)²) / n]
The mean square error measures the average of the square of the error. Because of the fact we use
the average forecasts, we do not sum up the squared error divided by the number and do not correct
for n, but we end up having one MSE value for each survey number. Therefore, in this study the
formula used is:
MSE = (forecasted – actual)²
B. Mean Absolute Error
MAE = (1/n) * ∑ |e|= (1/n) * ∑ |f - y|
The mean absolute error is the average of the absolute errors. As is the case with MSE, we rearrange
the formula to be able to compare the outcomes per survey number:
MAE = |error|

C. Mean Percentage Error

MPE = (1/n) * ∑ *(f – y)/y]
The mean percentage error is the computed average of the percentage errors by which the
predictions differ from the actual AEX values. Again, the formula we use is does not include the
summation sign and correction for n:
MPE = [(f – y)/y]

Because the respondents from our survey are anonymous, we are unfortunately not able to track
individuals throughout the entire survey period. The fact that the respondents are anonymous has
consequences for our calculations. To calculate the MSE, MAE and the MPE we take the average
forecasted values per survey number, instead of the forecasted value per individual. By using the

11
average values, we obtain a total of 12 values for the three methods. To compare the outcomes of
both scenarios, the average MSE, MAE and MPE is calculated of the 12 values.

For the MSE as well as the MAE and MPE holds; the smaller the outcome, the better because this
means the smaller are the differences between the predicted and the actual values. The final
outcomes can be compared at the end to see which of the two scenarios is more accurate in
predicting the AEX value correctly; the respondents with their predictions or the traditional random
walk approach? The same method is applied to compare the respondents’ predictions with the value
they give on what others will predict. Furthermore, we can see if there is a learning effect whenever
the values of MSE, MAE and MPE are decreasing throughout the survey period.

As pointed out before, the random walk model finds great support in literature on stock prices. In our
study, we look at to what extent the ABN Amro active clients are correct in predicting the AEX value.
Regarding expectations on the outcome of this study, we expect that the random walk approach
finds more support than the predictions made by the respondents. This means that in at least two
out of the three measures used, the scenario based on the random walk model needs to obtain
lower values. Our expectations can partly be explained by the fact that ABN recently started asking
their clients to make predictions, and that our data is rather limited because we have an average of
133 respondents. We can therefore suggest that the respondents have relatively little experience in
making predictions on the stock market, in this manner suggesting that the random walk approach is
more accurate in our case.
In addition to comparing the respondents’ predictions with the random walk approach, a second
comparison contains the values that respondents give each other. In this case we expect that
respondents consider themselves to be better predictors than the other respondents. In order to
support this view, the values of MSE, MAE and MPE need to be lower for the respondents’ predicted
values than for scenario in which the values are used of what respondents think that others will
predict.

12
5 Results
In this chapter, the results of this study are discussed. Three formulas are applied to two
comparisons, which will be discussed separately. Firstly, the comparison between the respondents’
predictions and the random walk approach is presented. Afterwards, the comparison is discussed
between the respondents and the values of what the respondents think others will predict.

5.1 Using predicted values: case (i)

In the first scenario the predicted values are used, which are given by the respondents through the
surveys. Firstly, for calculating the three formulas MSE, MAE and MPE the absolute error is
necessary, which describes the magnitude of the difference between the forecasted value and the
actual value. Per survey number, the average forecast value is used, since we deal with anonymous
respondents as explained earlier.
In the table on the next page, the absolute error is presented based on the average forecasted values
and the actual values on the forecast dates of the twelve survey numbers:

Forecast date 1. Average forecasted 2. Actual value Error = (1)-(2)

value
(1) Dec. 31 ‘09 326,456044 335,33 -8,87396
(2) Jan. 15 ‘10 335,6510417 337,99 -2,33896
(3) Jan. 29 ‘10 338,758427 327,9 10,85843
(4) Feb. 12 ‘10 330,5324675 315,74 14,79247
(5) Feb. 26 ‘10 317,3643411 317,74 -0,37566
(6) March 12 ‘10 319,3474576 339,57 -20,2225
(7) March 26 ‘10 341,3 343,81 -2,51
(8) April 9 ‘10 345,976378 355,89 -9,91362
(9) April 23 ‘10 355,4310345 353,38 2,051034
(10)May 7 ‘10 356,8349515 312,35 44,48495
(11)May 21 ‘10 320,7471264 313,41 7,337126
(12)June 4 ‘10 318,8024691 321,22 -2,41753

Table 1: presenting per survey number the absolute error ei = fi − yi, where fi is the prediction and yi the true value. The
prediction is in this case the predicted value by the respondents.

The average error in scenario 1 is 10.514695. This means that on average, the predicted value
differed 10.5 points from the actual AEX value.

5
Based on the absolute values of the errors in table 1.

13
After calculating the absolute errors of the survey number, the mean squared error, mean absolute
error and the mean percentage error can be calculated.

Forecast date MSE* MAE** MPE***

(1) Dec. 31 ‘09 78,7471 8,873956 -0,026463353
(2) Jan. 15 ‘10 5,470726 2,338958 -0,0069202
(3) Jan. 29 ‘10 117,9054 10,85843 0,033115056
(4) Feb. 12 ‘10 218,8171 14,79247 0,046850154
(5) Feb. 26 ‘10 0,14112 0,375659 -0,001182284
(6) March 12 ‘10 408,9512 20,22254 -0,059553383
(7) March 26 ‘10 6,3001 2,51 -0,007300544
(8) April 9 ‘10 98,2799 9,913622 -0,02785586
(9) April 23 ‘10 4,206742 2,051034 0,005804048
(10)May 7 ‘10 1978,911 44,48495 0,142420206
(11)May 21 ‘10 53,83342 7,337126 0,023410633
(12)June 4 ‘10 5,844455 2,417531 -0,007526091
Table 2: Using the predicted values  presenting outcomes of calculating values using the following formulas:
*MSE = (error)², **MAE =|error|, ***MPE = (error)/y

The results of the mean squared error in table 5 show that the MSE is in many cases below 1, despite
the outliers of survey numbers (6) and (10). The MAE is like the MSE (and the MPE) a negatively-
oriented score. The closer to zero, the better. In the results of scenario 1 all values are close to zero,
with some (for example (5) and (9)) more than others (for example (10)). The mean percentage error
describes the forecast accuracy as a percentage, with values close to zero indicating accurate
forecasting. The MPE values in our case are very close to zero, which suggests that the forecast were
accurate by the standard of this measure. Obviously a value of zero for all three methods is perfect,
suggesting that the predictions are perfectly accurate. However, in the stock market, as in many
other cases, making perfect predictions is only possible if one is clairvoyant.
When comparing the MSE outcomes of the first and last survey number, an improvement can be
observed, since the last value is significantly lower than the with the first survey number. We cannot
speak of a learning effect6, since the values of MSE differ greatly throughout the entire survey period.
The same conclusions can be applied to the MAE and the MPE; the value for survey number 12 is
smaller than for survey number 1, but we cannot mention a learning effect since the values differ to
a great extent throughout the survey period.

6
We assume a learning effect when the values for MSE, MAE and MPE are decreasing in time.

14
In order to be able to compare the outcomes of the different values for MSE, MAE and MPE, the
average values are calculated. After completing the same procedure for scenario (ii), a comparison
can be made. The scenario that contains the lowest values for MSE, MAE and MPE, is regarded to be
best in explaining the variability in observations in making predictions concerning the value of the
AEX index.

Average values for case (i), based on predicted values:

MSE= 248,1174 MAE= 10,51469 MPE= 0,011120407

5.2 Using communicated values: case (ii)

In the second scenario, we base our calculations on the well-known random walk model. Instead of
using the predicted values of the respondents, the communicated value is used. The value
communicated to the respondents includes the actual AEX value on the date of the previous survey
number. Therefore, in scenario two we assume a zero change in the AEX value. Following the same
method as used in scenario 1, we firstly obtain the absolute errors:

Forecast date 1.Communicated 2. Actual value Error = (1)-(2)

value
(1) Dec. 31 ‘09 324,63 335,33 -10,7
(2) Jan. 15 ‘10 335,33 337,99 -2,66
(3) Jan. 29 ‘10 337,99 327,9 10,09
(4) Feb. 12 ‘10 327,9 315,74 12,16
(5) Feb. 26 ‘10 315,74 317,74 -2
(6) March 12 ‘10 317,74 339,57 -21,83
(7) March 26 ‘10 339,57 343,81 -4,24
(8) April 9 ‘10 343,81 355,89 -12,08
(9) April 23 ‘10 355,89 353,38 2,51
(10)May 7 ‘10 353,38 312,35 41,03
(11)May 21 ‘10 312,35 313,41 -1,06
(12)June 4 ‘10 313,41 321,22 -7,81
Table 3: presenting per survey number the absolute error ei = fi − yi, where fi is the prediction and yi the true value. The
prediction in this case is the communicated value.

The average absolute error in scenario two is given by 10.680837, which means that on average, the
AEX value differed 10.7 points with the communicated AEX value. Subsequently, the three formulas
can be applied:

7
Based on the absolute values of the errors in Table 3.

15
 Forecast date MSE* MAE** MPE***
(1) Dec. 31 ‘09 114,49 10,7 -0,031908866
(2) Jan. 15 ‘10 7,0756 2,66 -0,007870055
(3) Jan. 29 ‘10 101,8081 10,09 0,030771577
(4) Feb. 12 ‘10 147,8656 12,16 0,0385127
(5) Feb. 26 ‘10 4 2 -0,006294455
(6) March 12 ‘10 476,5489 21,83 -0,064287187
(7) March 26 ‘10 17,9776 4,24 -0,012332393
(8) April 9 ‘10 145,9264 12,08 -0,033943072
(9) April 23 ‘10 6,3001 2,51 0,007102835
(10)May 7 ‘10 1683,461 41,03 0,131359052
(11)May 21 ‘10 1,1236 1,06 -0,003382151
(12)June 4 ‘10 60,9961 7,81 -0,024313555
Table 4: presenting outcomes of calculating formulas:
*MSE = (forecasted – actual)², **MAE =|error|, ***MPE = [(f – y)/y]

For the values of the MSE, we see that the value differs greatly throughout the survey period. The
value of the last survey is smaller than from the first survey, but a pattern of a decreasing MSE in
time cannot be observed. In the case of the MAE, the values do not differ that significantly as with
the MSE, but neither a learning effect can be observed. The similar story holds for the MPE, which
does not fluctuate enormously but does not show a decreasing value.
Yet again as in scenario 1, the average values of the MSE, MAE and MPE are needed to make the
comparison at the end.

Average values for case (ii), based on the communicated value:

MSE= 230,6311 MAE= 10,68083 MPE= 0,004338908

5.3 Comparing cases (i) and (ii)

In the concluding table, the average values for MSE, MAE and MPE are presented to make a
comparison.

Scenario Average MSE Average MAE Average MPE

1 (predicted values) 248,1174 10,51469 0,011120407
2 (communicated 230,6311 10,68083 0,004338908
value)
Table 5: presenting outcomes of calculating formulas with two scenarios:
(1) Using the predicted values by the respondents
(2) Using the communicated value

16
After using three measures to investigate which scenario makes better predictions concerning the
AEX value, it can be concluded that scenario 2 finds more support. When analyzing the measures
separately, the following conclusions can be made:

The mean squared error: scenario 2 contains a lower value, to be exact 7.6 % lower than scenario 1.
The mean absolute error: scenario 1 has a lower value for the MAE than scenario 2, with a difference
of 1.6 % compared to scenario 1.
The mean percentage error: finally, the last method also shows a lower value in case of scenario 2.
The average MPE using the communicated value is 70.0 % lower compared with the use of predicted
values.

We expected that using the communicated value, based on the random walk model, would find more
support than the predictors of the ABN Amro panel. We investigated this hypothesis by using three
mean square methods. In indeed two out of the three methods, the random walk model acquires
more support than the predicted values from ABN Amro clients. Our findings indicate that the
traditional random walk model is in the case of predicting the AEX value more accurate than the
predictions made by ABN Amro clients. Unfortunately our data is too limited to investigate whether
or not the AEX values follow a random walk, but we can assume that using the principles of the
random walk model leads to more accurate findings than in the case that the ABN Amro clients make
predictions on their own. For the ABN investors making predictions on the AEX value, it could
therefore be suggested to look at the last previous value instead of basing the prediction on other
values that include their sentiment.

5.4 From the respondents’ perspective: what will others predict?

Besides comparing the respondents’ predictions with the random walk approach, another
comparison of interest includes the expected AEX value that respondents give to other respondents.
In the questionnaire respondents were asked to fill in both a prediction for themselves, as well as
predictions for others. Though we assume that respondents consider themselves to predict more
accurately, we will observe that this assumption is not supported in our study.

To make calculations using the same method as before, we start with the error:

17
 Forecast date 1.Average predicted 2. Actual value Error = (1)-(2)
value
(13)Dec. 31 ‘09 326,4615385 335,33 -8,868461538
(14)Jan. 15 ‘10 338,0729167 337,99 0,082916667
(15)Jan. 29 ‘10 339,6235955 327,9 11,72359551
(16)Feb. 12 ‘10 330,6883117 315,74 14,94831169
(17)Feb. 26 ‘10 317,5503876 317,74 -0,189612403
(18)March 12 ‘10 318,4661017 339,57 -21,10389831
(19)March 26 ‘10 341,7076923 343,81 -2,102307692
(20)April 9 ‘10 345,6062992 355,89 -10,28370079
(21)April 23 ‘10 355,7931034 353,38 2,413103448
(22)May 7 ‘10 356,3883495 312,35 44,03834951
(23)May 21 ‘10 315,4137931 313,41 2,003793103
(24)June 4 ‘10 315,5925926 321,22 -5,627407407
Table 6: presenting per survey number the absolute error ei = fi − yi, where fi is the prediction and yi the true value. The
prediction in this case is the communicated value.

The average absolute difference between the predicted value and the actual is 10.2821215, which
means that on average the predicted AEX value differs 10.28 points from the actual AEX value. This
difference is lower than in the case of the respondents themselves, which is 10.5 points.
Continuing with the remaining calculations, we present hereby the outcomes of the MSE, MAE and
the MPE.

Forecast date MSE* MAE** MPE***

(1) Dec. 31 ‘09 78,64961006 8,86846154 -0,026446967
(2) Jan. 15 ‘10 0,006875174 0,08291667 0,000245323
(3) Jan. 29 ‘10 137,4426916 11,7235955 0,03575357
(4) Feb. 12 ‘10 223,4520223 14,9483117 0,047343738
(5) Feb. 26 ‘10 0,035952863 0,1896124 -0,000596753
(6) March 12 ‘10 445,3745237 21,1038983 -0,062148889
(7) March 26 ‘10 4,419697633 2,10230769 -0,006114737
(8) April 9 ‘10 105,7545019 10,2837008 -0,028895728
5,823068252 2,41310345 0,006828636
(9) April 23 ‘10
1939,376228 44,0383495 0,140990394
(10)May 7 ‘10
4,015186801 2,0037931 0,00639352
(11)May 21 ‘10
31,66771413 5,62740741 -0,017518858
(12)June 4 ‘10
Table 7: presenting outcomes of calculating formulas:
*MSE = (forecasted – actual)², **MAE =|error|, ***MPE = [(f – y)/y]

The values of MSE, MAE and MPE are not decreasing in time, even though the values for survey
number 1 < survey number 12. Therefore, we cannot notice an obvious learning effect, just like in
case as of the respondents’ predictions.
Proceeding the comparison, we can evaluate the outcomes with scenario 1 (using the predictions of
the respondents):

18
Scenario Average MSE Average MAE Average MPE
1 (predicted values) 248,1174 10,51469 0,011120407
3 (what others will do) 248,001506 10,2821215 0,007986104

Table 8: presenting outcomes of calculating formulas with two scenarios:

(1) Using the predicted values by the respondents
(3) Using the what respondents think others will predict

The table above shows that for the MSE, MAE and MPE all values are better for scenario 3. These
results suggest that respondents of the questionnaire make more accurate predictions for others
than for themselves. To be precisely per method:

The mean squared error: scenario 3 contains a lower value; 4.4 % lower than scenario 1.
The mean absolute error: scenario 3 includes a lower value than scenario 1, with a difference of 2.2%
compared to scenario 1.
The mean percentage error: also the last method shows a lower value in case of scenario 3. The
average MPE in scenario 3 is 28.2% lower compared with the use of predicted values.

It can be assumed that respondents consider themselves to be risk-seeking if they compare

themselves to other investors, in so doing willing to take more risk than what other investors would
take. However, we cannot make statements on this subject for the reason that we deal with
anonymous respondents of the survey. It can be concluded however that the respondents of our
survey appear overconfidence when it comes to predicting the AEX value; the confidence in
predicting the correct AEX value is greater than the actual accuracy of the prediction.
The recommendation conform our results is that when one wants to predict most accurately the AEX
value, one should not make their own predictions, but should predict that they think others will
predict.
In conclusion, we can say that in both the comparisons made in this study, the approaches based on
the literature and based on what others will predict, present more accurate results than in the case
of using the respondents’ predictions. Therefore, support is given to the traditional random walk
method and the respondents show overconfidence towards predicting. Our results confirm that
making predictions on the AEX value contains more accurate values whenever one predicts the value
that one thinks other investors will predict. What is confirmed first and foremost in this study is that
making predictions on the stock market is a difficult task and involves many factors that influence
decision making.

19
6 Conclusion and Discussion

Making predictions on the stock market involves many factors and is therefore of great interest in
the financial literature. Despite the support for the traditional view that the stock market is efficient
and follows a random walk, more room is made for behavioral economics in the sense that sentiment
becomes more prominent in investment decision-making.
In this study a comparison is made between on the one hand predictions made by ABN Amro clients
(scenario 1), and on the other hand the ‘traditional’ random walk model (scenario 2). Assumptions
from the random walk model include a zero change in the AEX value, therefore the value of the AEX
prior to the survey date is used which is communicated to the respondents. In addition, a second
comparison is made between the predictions of respondents and the scenario that uses the expected
AEX values that respondents give to other respondents (scenario 3).
Three methods are used to make the comparisons; the mean squared error MSE, the mean absolute
error MAE and the mean percentage error MPE. To interpret the results, it counts for all methods
that the smallest values indicate more accurate results and are therefore preferred.
Based on our results, in both comparisons the approach based on the literature delivers more
accurate results. In both comparisons the scenarios based on the random walk model (2) and on
what is thought that others will predict (3) comprised more accurate results in at least two out the
three methods MSE, MAE and MPE.
It can be suggested that in the case of ABN Amro clients giving expected AEX values, using the
communicated value gives more accurate results than making predictions on itself. The use of the
random walk model in this study does not implicate that the AEX value follows a random walk, for
the reason that we cannot make suchlike assumptions based on our data. In our comparison, the
application of the random walk model basically generates more accurate outcomes than applying the
predicted values. Furthermore, for the respondents to obtain more correct predictions, it can be
recommended to predict the same value as what they think other respondents will predict.
Based on the results of both comparisons, it can be stated that the predictions made by the
respondents do not provide the most accurate results. We investigated that by taking either the
actual AEX value of the previous survey date or by taking the value the respondents think that others
will predict, the respondents would make more accurate predictions on the AEX value than when
they use their own prediction.
More investigation could be done with the provided dataset when it would be for example possible
to follow respondents throughout the survey period. Furthermore, a greater number of respondents
would lead to stronger results and more possibilities in research.

20
7 References

Baker, M. and Wurgler, J. (2007) “Investment sentiment in the stock market”

Bikhchandani, S. and Sharma, S. (2000) “Herd behavior in Financial Markets: a review”

Black, F. (1986) “Noise”

Bradford DeLong, J., Shleifer, A., Summers, L.H., Waldmann, R.J. (1990) “Noise trader risk in financial
markets”

Brown, G. W. and Cliff, M.T. (2004) "Investor Sentiment and the Near-Term Stock Market"

De Bondt, W.F.M. (1985) “Does the Stock Market Overreact on New information?”

Dubois, D. and Prade, H. (1991) “Fuzzy sets in approximate reasoning, Part 1: inference with
possibility distributions”

Kahneman, D. and Tversky, A. (1982) “Intuitive prediction: biases and corrective procedures”

Keynes, J.M. (1964, reprint of the 1936 edition), p 153-154) “The General Theory of Employment,
Interest and Money”

Lehmann, B.N. (1990) “Fads, martingales, and market efficiency”

Malkiel, B.G.( 2003) “The efficient market hypothesis and its critics”

Malkiel, B.G. (2004) “Models of stock market predictability”

Morck, R. Shleifer, A. Vishny, R.W. (1990) “The Stock Market and Investment: is the Market a
Sideshow?”

Shleifer, A. and Summers, L.H. (1990) “The noise trader approach to finance”

Solt, M.E., Statman, E. (1998) “How useful is the sentiment index?”

Qiu, LilyX ., and Ivo Welch. (2006) "Investor Sentiment Measures."

Wolski, R. (1998) “Dynamically forecasting network performance using the network weather service”

21
8 Appendix

22