Investors’ cognitive bias and sell-side analyst information

(1)

1

Investors’ cognitive bias and

sell-side analyst information

Bachelor Thesis – Economie en Bedrijfskunde

(Accountancy & Control)

Gijs de Bra

11036168

Supervisor: Máté Széles

23 June 2018

Amsterdam Business School, Faculty of Economics & Business

Content:

(2)

Verklaring eigen werk

Hierbij verklaar ik, Gijs de Bra, dat ik deze scriptie zelf geschreven heb en dat ik de volledige verantwoordelijkheid op me neem voor de inhoud ervan.

Ik bevestig dat de tekst en het werk dat in deze scriptie gepresenteerd wordt origineel is en dat ik geen gebruik heb gemaakt van andere bronnen dan die welke in de tekst en in de referenties worden genoemd.

De Faculteit Economie en Bedrijfskunde is alleen verantwoordelijk voor de begeleiding tot het inleveren van de scriptie, niet voor de inhoud.

(3)

Abstract

The idea of rational behaviour by economic agents has been challenged in prior literature since the introduction of psychological biases and heuristics. This approach has also been adopted by finance researchers. Novel theories provide salient descriptions of investor behaviour, inconsistent with traditional models. However, one anomaly from accounting research remains generally disregarded. Therefore, I examine whether cognitive biases explain investors’ reliance on sell-side analyst information. I empirically test the hypotheses using market data of the largest publicly listed firms in the US. Specifically, I analyse volume and price responses to publications of analyst information. I find statistical evidence for one of two selected biases, but more importantly, I present supplemental data with interesting insights that signal presence of both. I infer that the results are provisionally positive and may be corroborated by further research. I contribute to finance research by expanding the application of cognitive biases. I also contribute to accounting research by examining how analyst information is processed and used for decision making by investors if they are affected by cognitive biases. Investors and regulators may also find the results useful for practice.

(4)

Samenvatting

Het idee dat economische agenten zich rationeel gedragen wordt sinds de vorige eeuw bestreden. Dit proces is in gang gezet door psychologische onderzoekers die cognitieve fouten en vuistregels ontdekten in gedragsexperimenten. De hierdoor ontwikkelde theorieën zijn vervolgens ook in de financiële wetenschap toegepast, waardoor het gedrag van investeerders beter verklaard kan worden dan met traditionele modellen. Een verschijnsel dat nog niet op deze manier is onderzocht, doet zich voor in de accountingwetenschap. Het betreft analisten, die aan de verkoopkant van de aandelenmarkt informatie verspreiden over bedrijven, winsten voorspellen en aanbevelingen geven voor investeringen. Er wordt onderzocht of cognitieve fouten verklaren waarom investeerders significant reageren op de soorten informatie die zulke analisten verspreiden. De afgeleide hypotheses worden met behulp van marktdata getoetst, waarbij de focus ligt op grotere bedrijven. De invloed van analisteninformatie op handelsvolumes en aandelenprijzen staat daarbij centraal. De resultaten zijn wisselvallig maar neigen naar bevestiging van de verwachtingen. In eerste instantie wordt voor één van de twee geselecteerde cognitieve fouten statistisch bewijs gevonden. Echter, uit analyse van additionele data blijkt dat beide fouten op zijn minst enigszins van invloed zijn. Derhalve kan op basis van de resultaten en de beperkingen van de methodologie enkel een voorlopig bevestigend antwoord gegeven worden op de onderzoeksvraag. Er zijn echter mogelijkheden voor vervolgonderzoek waaruit wel definitieve antwoorden kunnen voortkomen. Met het onderzoek wordt aan de financiële wetenschap bijgedragen middels uitbreiding van het toepassingsgebied van cognitieve effecten. De bijdrage aan de accountingwetenschap is dat er onderzocht wordt hoe de besluitvorming van investeerders op basis van analisteninformatie beïnvloed wordt door cognitieve fouten. Deze exogene variabele wordt immers vaak buiten beschouwing gelaten bij onderzoek naar informatieverwerking. Tenslotte kunnen toezichthouders met de resultaten het speelveld van analisten evalueren, terwijl investeerders hun beslissingen kunnen beoordelen en verbeteren.

(5)

List of Tables and Figures

Table Page

1 Recommendation revisions sample sizes in bull and bear markets 22

2 Target price publications sample selection 25

3 Descriptive statistics for the recommendation revisions samples 28

4 Correlations and descriptive statistics for the target price anchoring model 33

5 Multivariate regression output of target price anchoring model 34

6 Average abnormal returns for different implied return intervals 36

B1 Univariate regression model for the recommendation revisions samples 44

Figure Page

1 Frequency diagrams of the RRRs for all recommendation revisions samples 30

A1 Development of the monetary value of the S&P 500 index during the period of 44 January 2010 up to May 2018.

(7)

1 Introduction

Stock markets are interesting phenomena for many reasons. Not only are prices subject to a substantial amount of factors, but also investors are particularly unpredictable. Reasons for such dispersion in behaviour can be found by looking at personal characteristics. For instance, institutional investors may be risk-seeking and overconfident, while private investors may prefer safe bets so as not to jeopardise their life savings. Since it is evident that different types of investors have different motives and incentives, it seems implausible to assume that all of these economic agents make decisions in the same, rational manner. This is why in the past few decades researchers have looked for and found patterns in investor behaviour that are better explained by psychological theory than by economic theory (Hirshleifer, 2001). Such a transition has far-reaching implications for our understanding of how investors process information.

One particularly interesting type of information that investors incorporate in their investment decisions is provided by sell-side analysts. These information intermediaries follow publicly traded companies closely, publish earnings forecasts and provide recommendations to buy, sell or hold stocks. Because research finds that investors are highly sensitive to whether a company beats those forecasts or not (e.g., Bartov, Givoly, & Hayn, 2002) and that recommendation revisions lead to significant market responses as well (Jegadeesh & Kim, 2006), it can be argued that investors heavily rely on information that analysts disseminate. Economic theory predicts that investors use all information they gather and make informed decisions. However, in the first paragraph I present reasons why investor behaviour is not perfectly explained by traditional models. Therefore, psychological theory might provide better insights on investors’ dependency on analyst information.

The psychological approach to economic theory is also known as behavioural economics. This research field focuses on identifying patterns in behaviour of economic agents that are inconsistent with the assumptions of rational decision making. Important research in this field is done by Tversky and Kahneman (1974), who present an array of heuristics and biases that they find in experiments and conclude that these “lead to systematic and predictable errors” (p. 1131). Plenty of research exists on the wide range of subjects and settings these findings can be applied to. Since these cognitive biases are also extended to financial decision making, they can help solve anomalies of investor behaviour, such as underlying reasons for relying on analyst information. Therefore, I examine whether investors are subject to cognitive biases when they process information from sell-side analysts.

Prior research on analysts focuses on topics such as the biasedness of their earnings forecasts (Bissessur & Veenman, 2016), the value of analyst coverage (Li & You, 2015) and

(8)

market responses to earnings surprises (Bartov et al., 2002). However, most of this research disregards one side of the interaction between analysts and investors by designating analyst behaviour as the key factor in stock market responses. Nevertheless, disseminating information is only part of the transaction. The recipient still has to determine whether to use it or not and how to interpret it. A perfect example of such analyst-focused research is set by Jegadeesh and Kim (2006). They state that a stronger market response to recommendation revisions in the US than in other well-developed countries means that analysts from the United States (US) are more skilled, without considering investor characteristics that may differ across countries.

In order to determine whether cognitive biases dictate investor behaviour when analyst information is processed, this study focuses on several biases that are obtained from psychological theory. I empirically test their presence in the context outlined above with data of market prices as well as trading volumes. The selection depends on the type of bias and the related hypothesis. I use data of the different types of analyst information to explain those market movements, while taking cognitive biases into account. I largely control for other influences by narrowing down the circumstances, which increases the measurability of the biases.

The cognitive biases that I select are the confirmation bias and the anchoring bias. I find that the confirmation bias does not explain market responses to recommendation revisions. However, the distribution of responses provides insights which indicate that the bias affects investors, which creates an opportunity for future research. I also find that the anchoring bias explains market returns with respect to target prices. The positive relationship conforms to prior literature that documents the bias in other domains in a similar manner. With these results, I provide a provisional positive answer to the research question , which may be corroborated by further research.

I contribute to finance research by expanding the application of cognitive biases to analyst information. Furthermore, I contribute to accounting research by examining how information processing by investors is influenced by biases. Rather than documenting a significant response to analyst information, I analyse how the underlying reason for that response by incorporating behavioural factors. Additionally, the results indicate that investors unconsciously delegate their investment decisions to analysts. Such a conclusion is useful for investors because it helps them realise they are not making fully informed decisions, which negatively affects their returns (Mikhail, Walther, & Willis, 2007). It also tells regulators that investors rely on analyst information due to cognitive biases, which can be harmful to them. This may induce more supervision of analysts from the relevant authorities to prevent abuse.

Section 2 reviews literature on the three main themes of this study: investor behaviour, sell-side analyst information and cognitive biases. It also develops hypotheses based on the

(9)

integration of the matter. After that, section 3 outlines the methodology. Section 4 subsequently discusses the results and a conclusion is presented in section 5.

2 Literature review

This section analyses literature on the three central themes of this study. First, I discuss behavioural economics theory. Investor psychology is discussed in subsection 2. In the third subsection I examine literature on sell-side analysts so as to evaluate their performance as an information intermediary in stock markets. To conclude, I integrate all themes in subsection 4, in which I also analyse their relation and derive hypotheses.

2.1 Exploration of behavioural economics research

In the previous century, psychological research emerged that proposed alternative explanations of human behaviour in decision making. The contemporaneous standard in economic theory was rational decision making. This traditional model states that individuals make consistent, optimal decisions based on consideration of all possible choices. Simon (1957, as cited in Bazerman & Moore, 2017) challenges this model by hypothesising that decisions are not made in a fully rational manner, because the human mind has inevitable limits. Moreover, instead of endeavouring to optimise, people settle for satisfaction (Bazerman & Moore, 2017). These limitations to rationality, as illustrated in this subsection, lead to inconsistent and uninformed decisions.

The deviation from the assumption of rationality is facilitated by researchers that focus on heuristics and biases. Heuristics are described as “simplifying strategies, or rules of thumb” (Bazerman & Moore, 2017, p. 6). Some of the first experiments in this context were carried out by Tversky and Kahneman (1974). In their article, they summarise several experiments that result in the formulation of a specific heuristic and/or bias. For instance, they observe that participants rely on a predetermined value when estimating another value. These individuals do not adjust their estimation sufficiently, so that it is directed towards the initial value. Objective computation is no longer achieved, regardless of whether the initial value is relevant or not. Tversky and Kahneman call this the anchoring bias (1974). A practical example of such an initial value is the listing price of residential homes, whereas the value to be estimated is the fundamental value of the home. Northcraft & Neale (1987) show that both experienced and inexperienced estimators implicitly rely on the listing price, even if they possess plenty of objective information. They state that such a result is interesting because it is a valid example of applying experimental findings (i.e., the heuristic or bias) to real-life circumstances.

(10)

Other notable cases are the confirmation bias, the representativeness heuristic and overconfidence. The confirmation bias causes individuals to look for and accept information that confirms their beliefs or hypotheses, while they ignore and criticise any information that disconfirms them (Kunda, 1990). Individuals that are subject to the confirmation bias significantly underestimate the amount of debate that exists about their beliefs, so that they are hardly ever challenged. The representativeness heuristic is used when “the likelihood of an event is evaluated by the degree to which it is representative of the major characteristics of the process or population from which it originated” (Tversky & Kahneman, 1972, p. 451). A salient example is the translation of population characteristics to samples. Although samples can approach normal distributions if they are sufficiently large, there is no guarantee that the marginal event in such a sample will resemble that distribution. However, individuals seem to believe so because they disregard the difference between sample and population (Tversky & Kahneman, 1972). Overconfidence is a generic personality trait that is also considered a decisional bias. It is defined as follows: “the tendency to overestimate the likelihood that one’s favored outcome will occur” (Griffin & Varey, 1996, p. 228). This implies that individuals make decisions that rational agents would not, because information is distorted from reality in the process. Overconfidence is widespread and leads to judgmental errors, so it should be accounted for in behavioural models, according to Griffin and Varey (1996). Besides these biases, there are many more examples in research. For an in-depth discussion of them, I refer to Bazerman and Moore (2017) or Hirshleifer (2001).

A tangent stream within this psychological field specifically focuses on decisions that are related to risk. The prospect theory, introduced by Kahneman and Tversky (1979), challenges expected utility theory in explaining behaviour of economic agents. It portrays several characteristics which are persistent patterns of behaviour, similarly to heuristics and biases. For instance, loss aversion causes individuals to consistently attach more weight to losses than to gains of an equal amount. These are considered incremental outcomes and are compared to a reference point, not the final state of wealth. This also results in dispersed risk preferences with respect to gains and losses; individuals tend to be more risk-seeking when facing a potential loss (Kahneman & Tversky, 1979). In the application of these results to financial decision making, evidence is found that investors are reluctant to sell stocks at a loss (thereby taking a risk to turn them into gains) while capitalising on opportunities for gains as quickly as possible. This tendency, described by Odean (1998), is known as the disposition effect. It is one of the early findings in the combined field of psychological science and finance, known as behavioural finance, which is elaborated in the next subsection.

(11)

2.2 Investor behaviour in stock markets

In traditional finance textbooks, markets are assumed to be efficient, which simplifies theory relating to project and security valuation. An efficient market is defined as a competitive environment in which market prices are determined by “all information that is available to investors” (Berk & DeMarzo, 2014, p. 295). However, there is plenty of research on stock markets that shows that the aforementioned assumption of the efficient market hypothesis (EMH) is rarely sustained in practice. Violations can generally be assigned to two categories. The first one is the existence of private information. This includes information that is difficult to find or to interpret (so while it is publicly available), which limits the amount of investors that are able to use it (Berk & DeMarzo, 2014). The second category refers to investors having access to a large amount of information but not using all of it (objectively) for their investment decisions. This subsection integrates research that opposes market efficiency and rational decision making in financial contexts.

Behavioural finance research attempts to explain various instances of the second violation. Its impact on asset pricing is examined by Hirshleifer (2001), who surveys an array of psychological effects on stock markets. He concludes that investors make errors in assessing the value of assets, which is unpredicted by rational theory. By translating these errors into more coherent theories, researchers have initiated a transition from rational models and EMH to behavioural models. In general terms, this means that investor behaviour is no longer predicted and prescribed, as is the case for traditional rational models, but described and explained. Research empirically supports the notion that investors deviate from rationality. Therefore, descriptive information may help them improve their future decisions, while prescriptive models inhibit learning from mistakes that result from using them (Bazerman & Moore, 2017). This is because they do not adapt to those mistakes and assume rationality. I discuss research that attempts to find descriptive theories about investors below.

A frequently cited article in this context by De Bondt and Thaler (1985) provides empirical evidence for overreaction on stock markets. It illustrates a strong initial response to news that is followed by a reverse price movement in subsequent periods. The conclusion implies that investors overestimate the value of any information at first and they correct for it at a later stage, or that some investors overreact and others profit at their expense by correcting for it. Such patterns of behaviour are explained by Daniel et al. (1998), who model investors’ overreaction to private information in particular. They state that such market movements are a result of overconfidence, which causes investors to underestimate risk, underreact to public news and “lose money on average” (Daniel et al., 1998, p. 1866). It should be noted that their model is not tested for predictability in real-life circumstances. Nevertheless,

(12)

the theory approaches the desired function of a descriptive model. This is because it tells investors where the problem lies, instead of merely reminding them of its presence.

Besides market-wide psychology, cognitive biases that relate to specific pieces of information are also examined. For example, George and Hwang (2004) show that short-term returns are correlated with the extent to which a stock price approaches a historical price benchmark (the 52-week high), because it is an indication of recent good news. This result supports the hypothesis that the benchmark serves as an anchor that influences investors’ subjective valuation of the stock. Li and Yu (2012) find the same result, only with a different focus. The anchor they find is the 52-week high of the Dow Jones Index. Such a finding is interesting because it indicates that investors rely on more than just firm-specific anchors. Although it is regular that market information is incorporated into prices, a direct linear relation based on a cognitive bias is not straightforward.

Another bias that serves as an impediment to informativeness of market prices is the confirmation bias. When investors have access to a wide range of information, Chang and Cheng (2015) find that they simply focus on pieces that are in conformity with their personal beliefs. Investments decisions and prices are then biased towards these beliefs, at least in the short-term. They also find that this is especially the case for information that is not salient (qualitative, not related to earnings announcements, etc.). This result may explain why De Bondt and Thaler (1985) observe systematic overreaction to news: the only investors that react to news are the ones whose beliefs are confirmed, leading to an unmitigated response. Hirshleifer (2001) applies this bias to investors that “stick to unsuccessful trading strategies, causing mispricing to exist” (p. 1549). In short, receptions of as well as responses to information are dependent on investors’ beliefs or opinions.

While the biases describe investors’ tendency to focus on information consistently, Peng and Xiong (2006) find that such situations may change over time. They state that limited investor attention leads to a high amount of market-wide information processing. However, when investors increase their attention level, it induces a gradual transition to information about the firm that they have invested in. Such a process increases informativeness of prices, so that market efficiency increases. It paradoxically also leads to overreaction to news because investors overestimate their ability of processing firm-specific information (Peng & Xiong, 2006). Thus, an optimal mix of information is difficult to accomplish, but it may be beneficiary for investors to evaluate and adjust (if needed) their limited focus.

To conclude, the research field that finds these violations of market efficiency, also known as behavioural finance, examines several psychological biases. Evidence shows that their effects are generally credible, but further applications are yet to be explored. I examine whether cognitive biases explain investor behaviour when analyst information is processed, thereby contributing to the research field mentioned above.

(13)

2.3 Sell-side analysts and their role as information intermediary

To aid investors in making their investment decisions, sell-side analysts disseminate summarised information about publicly traded companies based on the research they conduct. In the process, they supposedly reduce information asymmetry. Some commonly known types of publications are quarterly or annual earnings forecasts, target prices and recommendations to sell, hold or buy stocks. Together they provide investors with relevant information regarding future price movements (Brav & Lehavy, 2003). In this subsection, I discuss analyst information in detail and evaluate the analyst as an information intermediary, so that the relationship with behavioural finance becomes more vivid.

First of all, it is useful to explain the content of analyst information before evaluating its value. Analyst recommendations are simple statements that summarise the analysts’ opinion about the stock’s future movements. They usually range from strong sell to strong buy. Target prices are expected future stock prices that analysts often include in their reports. They are quite straightforward as well, except that it is often not clear when exactly the analyst expects the price to reach its target (Brav & Lehavy, 2003).1_{The final type to be defined is the earnings}

forecast. Analysts attempt to predict what a company’s earnings per share (EPS) will be, usually for each quarter and fiscal year.

In order to facilitate investors’ decisions efficiently and effectively, these types of analyst information are expected to be free from bias. However, research shows that objectiveness of analyst information is far from guaranteed. For instance, Bissessur and Veenman (2016) find that analysts are incentivised by managers to revise EPS forecasts in a biased, downward direction so that positive earnings surprises (actual EPS minus forecasted EPS) are easier to achieve. This is because managers are aware of the benefits of beating EPS forecasts and the drawbacks of not meeting that objective, which they also try to achieve by managing earnings (Bhojraj, Hribar, Picconi, & McInnis, 2009). In addition, Veenman and Vermijmeren (2018) show that those pessimistically biased forecasts persist because investors do not entirely discern and appreciate their incentivised background. Such observations are remarkable because the roles are reversed. While analysts would normally endeavour to forecast firms’ EPS as accurately as possible, in this situation managers of those firms attempt to meet or beat analysts’ forecasts. This behaviour is explained by looking at the benefits for those firms. Bartov et al. (2002) find significantly higher stock returns for firms that have positive earnings surprises than firms who do not. These results also apply to firms whose managers engage in earnings management to beat analysts’ forecasts. The researchers also

1_{Gerritsen & Weitzel (2017, p. 1) state that a stock price is expected to approach the target price somewhere within}

(14)

find that such performance does not reverse in the long run. It is clear that firms have incentives to meet or beat forecasts and there is evidence that analysts help them doing so, so that forecasts are likely biased.

While investors respond strongly to earnings surprises, analysts have the tendency to underreact. Shane and Brous (2001) show that their underreaction causes analysts’ subsequent forecasts to remain biased until other information corrects it. More importantly, they conclude that such behaviour induces biased market responses too so that it is an irrational situation altogether. In addition, Ramnath (2002) finds that analysts underreact to EPS forecast errors for firms within the same industry. He illustrates this by comparing consecutive earnings surprises of those related firms, which provide a good benchmark, although firm-specific information as a factor is not appreciated. Nonetheless, such forecast errors may conceal relevant market information (such as abnormal demand levels) and thus analysts may be expected to learn from them. Therefore, analysts make errors persistently and fail to learn from them, so that they create another reason for bias in their reports.

Similarly to EPS forecasts, recommendations are found to provoke strong market responses. Especially institutional (as a proxy for large) investors make use of the information that such a recommendation revision provides. Smaller investors naively “tend to react more to the mere occurrence of a recommendation” (Mikhail et al., 2007, p. 1250) so that they always buy more than they sell. Therefore, it is comprehensible that Chen and Cheng (2006) find that institutional investors gain from trading on the basis of those revisions, implying predictable returns. It should be noted that the researchers model for as many other factors as possible and still find a low R2_{, indicating that recommendation revisions explain very little of the}

variance of institutional trades. Nonetheless, an interesting notion is that institutional investors may receive analyst reports earlier than the market does (Chen & Cheng, 2006). This is a perfect example of a first category violation of EMH’s information assumption, which relates to private information. To conclude, recommendation revisions are utilised by both large and small investors, but in different manners with negative consequences for smaller investors.

A particularly relevant article about the value of recommendations and international disparities is written by Jegadeesh and Kim (2006). They examine market responses to recommendation revisions in several wealthy countries and find significant responses in most of them. According to the researchers, the higher returns and response volumes in the US indicate a greater level of skill of US analysts. Paradoxically, they also find that recommendations by US analysts contain the strongest optimistic bias, so that investors are incentivised to buy stocks, more so than in other countries. Also, the researchers seem to disregard that investors’ (or economic agents’ in general) characteristics differ across countries. This is illustrated by Hofstede (1983), who quantifies them through questionnaires. He finds, for instance, that individuals from the US have higher levels of motivation and

(15)

self-centredness in comparison to other countries. This may influence how they respond to incentives and information. Therefore, the differences between market responses in different countries are quite possibly a result of intercultural discrepancies.

The nationality approach also applies to the rejection of an alternative hypothesis in the article. According to Jegadeesh and Kim (2006), the US market is not different (in terms of efficiency) from other countries, regardless of stronger responses to revisions on US stock recommendations from US analysts than from non-US analysts. They state that it just means US analysts are more skilled, leading to higher quality reports. This implies that market responses to US analysts’ reports are more rational. This is disputable for two reasons. First, US analysts may be better informed about US stocks, so their skill does not necessarily apply to stocks in other countries (Jegadeesh & Kim, 2006). Second, there is research that finds a home bias, i.e. investors are biased in their search for stock investments by prioritising stocks in their home country (Cooper & Kaplanis, 1994). Therefore, investors from other countries are less likely to buy US stocks, which may explain why the market responses to revisions on US stocks from non-US analysts are weaker (assuming non-US investors mainly follow analysts from their own country). The conclusion is that such research on analysts disregards the receiving end of information, deliberately or not, and thereby does not account for investor behaviour.

Besides discussing several biases in analyst information and related investor responses, I also find it necessary to determine the inherent value analysts add. First of all, since they are information intermediaries, analysts are expected to decrease the information asymmetry between companies and investors. Li and You (2015) find no statistical support for this notion with respect to company information, but they do find that trading volumes of stocks increase when analysts start following them. According to the researchers, this result indicates that more investors become knowledgeable about those companies, which is actually a decrease in information asymmetry after all. Another observation by Piotroski and Roulstone (2004) is that analyst coverage leads to more industry-related information incorporated in prices due to analysts’ industrial expertise, which explains synchronous price movement (similarly to Ramnath [2002]). While they state that such a relation indicates a more efficient incorporation of information, it also replaces firm-specific with market risk and return. Perhaps this is why less correlation with industry information is characterised as “equivalent to more informed prices” (Piotroski & Roulstone, 2004, p. 1147). Thus, analysts decrease information asymmetry by increasing awareness for companies they start following and by putting emphasis on industry information, although their added value has constraints due to biased information.

What I illustrate in this subsection is that research provides support for the idea that analysts publish biased information on which investors rely. If the information assumption of

(16)

EMH is reconsidered, there is a third violation possible: Some information that investors receive and process is biased and therefore not necessarily a fair representation of reality. This can be problematic due to the reliance. Referring to the second violation of the EMH assumption, it is important to examine how analyst information is used. One report can be interpreted differently by two investors, not to mention an entire stock market. I use cognitive biases from behavioural economics to explain this process and incorporate the existing evidence about inefficient investor behaviour.

2.4 Cognitive biases in an investor-analyst context

As I showed in the previous subsections, there is substantial research on investors that attempts to explain their information-processing behaviour. On the other side of this study, accounting research finds incentives that explain why analysts disseminate certain information to investors. However, the latter often does not appreciate that analyst information is interpreted in different ways so that its value to investors is either amplified or diminished. That is important to take into account, because investors’ limited attention influences the way they process information (Peng & Xiong, 2006). Limited attention is a common basis for many cognitive biases, because investors tend to focus on a specific piece of information. In this study, I examine whether the interpretation of analyst information by investors is affected by cognitive biases.

An example of a positive result in this context is presented by Balkanska (2018). She finds that more dispersion between analysts’ forecasts of a firm’s earnings is interpreted as greater uncertainty, similarly to Bissessur and Veenman (2016). Uncertainty is associated with risk, so that investors are inclined to sell winners (stocks with a gain) more quickly, while losers (stocks with a loss) are unaffected. This is because the disposition effect leads to “avoiding realization of losses” (Balkanska, 2018, p. 857). Investor mood as a cognitive factor is also applied to analyst information. Kliger and Kudryavtsev (2013) conclude that investor mood may explain significant responses to recommendations, since positive mood indicates positive expectations of the future and vice versa. Both of these articles use a psychological approach to analyst information, rather than a rational one.

Another instance relates to analysts’ target prices and mergers and acquisitions (M&A). According to Gerritsen and Weitzel (2017), stockholders of the selling firm compare takeover bids to analysts’ target prices, which then function as a reference point. Their results imply that takeover bids need to beat target prices in order to be interpreted as a gain by investors and succeed in the end. This effect is similar to the anchoring bias (Gerritsen & Weitzel, 2017), because it also implies that “different starting points yield different estimates” (Tversky & Kahneman, 1974, p. 1128), the starting point being the target price. That notion is applicable

(17)

to earnings forecasts as well, because investors respond more positively when firms beat them than when they do not (e.g., Bartov et al., 2002; Bhojraj et al., 2009). So, it seems possible that investors anchor on the occurrence of a positive earnings surprise, leading to a higher estimate of the stock value.

In the previous subsections, I discussed several biases and other investor characteristics. A comprehensive study could integrally test which are related to analyst information processing. In this study, however, the selection of biases to be examined is limited due to time constraints. I focus specifically on the confirmation bias and the anchoring bias, because they are testable, salient and applied to financial decision making before.2_{The latter}

is especially relevant for two reasons. First of all, applying experimental findings to situations outside the laboratory, in this context the stock market, improves the explanatory function of a theory or bias (Northcraft & Neale, 1987). Second, the multiplicity of extrapolations of a theory is considered a prerequisite for a behavioural theory to have predictive power (Daniel et al., 1998). Since the confirmation and the anchoring bias pass these tests, I consider them feasible.

Besides feasible, the selected biases are also interesting. The anchoring bias significantly influences estimates of any kind if there is an initial value that the individuals can rely on (Janiszewski & Uly, 2008). That value can be as irrelevant as possible and still function as such (Tversky & Kahneman, 1974). Also, since it is an implicit bias, even skilled estimators are affected (Northcraft & Neale, 1987). Therefore, unsophisticated and sophisticated investors may both anchor on any initial value in the stock market. The confirmation bias is not as tangible but still qualifies as interesting, because it applies to virtually every individual with beliefs. The bias represents opinion-based decision making, which is a more relatable and tangible concept. This is because beliefs are the basis of regular decisions, such as a political vote or choice of religion. The hypotheses related to these biases are derived below.

The first hypotheses regard the confirmation bias, which causes individuals to adhere to their beliefs. Since these are personal, it is essential in a market study to look for beliefs that may apply to a large amount of investors. A salient example is market sentiment, which is documented to affect prices significantly (DeBondt & Thaler, 1985). In general terms, the market can be said to be in two states of sentiment: bull or bear. A bull market indicates investor optimism and persistent price increases over a longer period of time (several years). A bear market refers to contrary circumstances. Therefore, a recommendation upgrade (e.g. ‘hold’ to ‘buy’) should confirm an investor’s optimistic beliefs in a bull market, while the opposite is presumed in the state of a bear market. I examine recommendation revisions instead of

2_{The anchoring bias applies to, for example, housing prices (Northcraft & Neale, 1987), historical stock or index}

benchmarks (Li & Yu, 2012) and target prices in M&A (Gerritsen & Weitzel, 2017). The confirmation bias applies to, for example, qualitative information (Chang & Cheng, 2015) and trading strategies (Hirshleifer, 2001).

(18)

earnings forecasts or target prices, because the information value of a revision is qualitative and therefore more comparable to qualitative beliefs such as positive sentiment. Quantitative information may be compared to other financial data, so that the focus on beliefs is diminished.3

Consequently, the following hypotheses are formed:

H1 Investors are subject to the confirmation bias when they interpret analysts’

recommendation upgrades in a bull market.

H2 Investors are subject to the confirmation bias when they interpret analysts’

recommendation downgrades in a bear market.

The second bias that I apply to analyst information is the anchoring bias. When testing this particular psychological effect, it is required to designate two values: the anchor and the value that the decision maker estimates. In this context, the most salient value to be estimated is that of a company’s stock, because such a valuation enables the investor to make trading decisions. There are several possible anchors on which investors may rely, such as the 52-week high (George & Hwang, 2004). I will focus on a vivid anchor that analysts provide, which is their personal expectation of the stock value, also known as the target price. Technically, there is a fundamental disparity between the two variables. Investors estimate the current value of a stock, so that they can compare it to the current market price. Analysts, on the contrary, implicitly estimate the future value of the stock, which is why it is labelled as a target. Therefore, a rational investor should not consider them directly comparable. However, there are reasons for investors to base their estimation on the target price. This future value is higher than the current price if, for example, analysts anticipate earnings that outperform expectations. Therefore, an investor that relies on analysts’ predictions may revise his expectations to match theirs. This in turn will lead to a higher stock value estimate. Furthermore, a target price that is higher than the current price can lead to positive returns in two ways. The first is the situation in which the analysts’ expectation is accurate and the price approaches the target price in the future. The second regards investors that anticipate other investors responding positively to the target price which contributes to an upward price movement. In conclusion, I consider the target price to be an anchor that influences investors’ estimate of the related stock value.

I could include earnings forecasts in the study, but market responses to them have been examined exuberantly. The results also relate more to whether forecasts are met, beaten

3_{For example, investors may be inclined to immediately compare a target price to the current stock price or a}

historical benchmark when they first observe it. Therefore, a different bias (e.g. anchoring) may be triggered first, leading to a market response straight away. Of course, market sentiment may impact that response, but in contrary to recommendation revisions, it is unlikely to be the main moderator variable.

(19)

or missed, rather than to an estimation process. Returning to target prices, there is empirical evidence that their publications are correlated with short-term returns (Brav & Lehavy, 2003). This implies that they play a role in the decision-making process of investors. Gerritsen and Weitzel (2017) find that this is the case in M&A negotiations. In this study, I hypothesise that the anchoring effect of the target price is of general presence in stock markets:

H3 Analysts’ target prices serve as a psychological anchor in investors’

valuation process of a stock.

The presence of an anchor is not the only relevant aspect of the bias. According to Janiszewski and Uly (2008), the precision of the anchor is relevant as well. Since the estimation process involves an adjustment away from that anchor, the researchers investigate whether the adjustment differs if the anchor is rounded, slightly lower than rounded or slightly higher. They find that the estimated values are indeed significantly different. A higher precision anchor (e.g., $199,900 is more precise than $200,000) decreases the adjustment, so the estimation is closer to the anchor than if the anchor is rounded (Janiszewski & Uly, 2008). In other words, less rounding increases the anchoring effect. I hypothesise that the target price functions similarly. A rounded price of $35.00 should induce a larger adjustment than a target price of $34.90 or $34.97. In other words:

H4 Analysts’ stock target prices that are less rounded stimulate an investor to value

the stock closer to them.

3 Data and sample selection

The challenge of empirically testing psychological effects with market data is that there is a substantial amount of exogenous factors. This makes it difficult to filter or prove behavioural effects. Also, if a correlation exists between a piece of information and a market response, it isn’t necessarily caused by a cognitive bias: It could be a rational decision. Third, since market responses are aggregates, it is debatable whether all participants are subject to the same bias. However, opportunities for measurement exist, for example by correcting for exogenous effects or adding additional variables that measure bias. Also, if a biased response is observed, it is more relevant that cognitive biases impact the market, rather than how many investors are subject to such biases.

There are arguments in favour of the credibility of empirically measuring a psychological bias if it “explains a range of empirical patterns in different contexts” (Hirshleifer,

(20)

2001, p. 1564). This implies that the measurability of a bias may increase if it has been empirically examined before. In the previous section, I discussed prior research that finds several instances of the two selected biases: anchoring and confirmation. Therefore, this supposed condition is not an impediment to the study. For both biases, two hypotheses were derived. The methodology for these hypotheses is outlined below.

3.1 Confirmation bias

The impact of the confirmation bias on the interpretation of analyst recommendation revisions is measured by analysing trading activity. First, I will discuss the selected samples. To distinguish between bull and bear markets, S&P 500 (as a market proxy) movements over time are analysed by Nyberg (2013, p. 3355). For instance, he shows bear markets between August 2000 and February 2003 and between October 2007 and February 2009. I include the months within those boundaries in the sample. I assume a bull market for the recent years due to positive market returns and a lack of economic turbulence.4_{Appendix A supports this notion}

and shows that the selected period (January 2012 - December 2017) matches the visual description of a bull market. In the nearest periods on the outside of those boundaries there is some volatility, which justifies the cut-offs.

The firm selection is as follows. Due to the time constraints of this study, I focus on the largest publicly traded firms. They are the most salient firms to investors and analysts and make up a large proportion of the stock market. More specifically, all companies in the S&P 100 index are included. This index distinguishes between firms on the basis of market capitalisation as a proxy for firm size. The selection of firms is obtained from Barchart.com (n.d.) on 21 May 2018 and included 102 stocks due to two firms having multiple share classes. I exclude one stock on the list, which is Ultra Semiconductors, because there is no analyst data related to it.5_{Large firms form a good selection for measuring the confirmation bias}

because they presumably provide the richest information environments possible. Those circumstances amplify the bias (Chang & Cheng, 2015). One limitation of this selection is that some firms in the current S&P 100 may not have been as relevant in the earlier periods. For example, Facebook did not exist during the 2000-2003 bear market. Therefore, in addition to the sample of firms being limited, its intertemporal consistency is constrained. This may also

4_{Nyberg (2013, p. 3355) also designates historical bull markets, but I use more recent data because advancing}

technology improves dissemination and reception of analyst information. Also, the most recent bull market he observes, starting in February 2009, experienced some volatility, as is shown in appendix A. The period after that appeared more consistent, except for the minor volatility around January 2016.

5_{This is because Ultra Semiconductors ProShares (official ticker: USD) is an Exchange-Traded Fund (ETF). This}

(21)

impede the applicability of results to other firms. Nonetheless, I keep the data selection concise, considering the time constraints as well as the preliminary nature of this study.

For the selected firms, recommendation revisions data are gathered from the I/B/E/S database. Recommendations are coded on a 1-5 scale, 1 being the highest recommendation (strong buy) and 5 the lowest (strong sell). This scale is reversed to make interpretation of the descriptive results easier. For the tests, however, the rating does not matter because it is only relevant whether the revision is an upgrade or a downgrade. The daily trading volumes are obtained from the CRSP database. Additional data is gathered from Nasdaq.com, which is necessary for some individual cases. This is because CRSP data are available up to 31 December 2017, so some trading volumes in January 2018 need to be gathered manually.

In Table 1 I summarise the amounts of recommendation revisions that the selected firms and periods yielded. Not all data were useful, because the identification of a revision requires a comparison to the previous recommendation by the same analyst. All recommendations by analyst codes of 0 are excluded, because it is uncertain whether they are published by the same person. Also, first recommendations within a period are excluded, which includes those by analysts who did not publish a second one. Furthermore, if a new recommendation is equal to the previous one, it is not a revision but a confirmation of prior information. The hypotheses require tests of revisions, even though these repetitions may be relevant. However, Mikhail et al. (2007) use a similar approach and find that including these “reiterations” (p. 1232) do not impact final outcomes. Furthermore, most of the exclusions were first and singular recommendations. I could have converted these to revisions by looking for the previous recommendations in earlier periods. It is therefore a limitation of this study, caused by time constraints, that not all published recommendations yield revision data. Nevertheless, the sample size is considerable and represents slightly less than a third of the total sample data.

The dependent variable is denoted as the recommendation revision response (RRR). I quantify this by calculating the difference between trading activity before and after the revision. Trading activity is represented by trading volumes, of which I compute the average per day to make the results comparable. So, the average of daily trading volumes (ATV) is calculated for those two periods. First, the trading volumes on the day of the revision and two days after the revision are included (t : t+2). I consider this a reasonable time frame for investors to respond to a revision, in conformity with prior research. Mikhail et al. (2007) incorporate two days after the revision, while Chen and Cheng (2006) find one day sufficient. Additionally, Kliger and Kudryavtsev (2013) only find significant returns within two days. Second, the measurement period before the revision comprises thirty trading days (t-30 : t-1). I incorporate such a large period before the revision because the average of it should represent a regular trading volume. This is based on the statistical premise that thirty observations or

(22)

Table 1

Recommendation revisions sample sizes in bull and bear markets

Bear market

Parts of sample Bull market 2000-2003 2007-2009 Total Initial number of recommendations 8,572 5,345 2,419 16,336 First and singular recommendations -5,969 -3,344 -1,808 -11,121 and recommendations equal to the

previous one

Recommendation revisions 2,580 2,000 610 5,190 Distribution:

Recommendation upgrades 1,373 848 305 2,526 Recommendation downgrades 1,207 1,152 305 2,664

Note. This table states the number of recommendation observations that are gathered from the I/B/E/S database. It also outlines how I derive my final sample size from the raw data amounts. The initial number of recommendations represents all data of the S&P 100 firms within the periods, excluding any observations from analysts whose code is 0. Revisions from first recommendations within the period are unidentifiable. Singular recommendations have no posterior observation that can be identified as a revision. Similarly, observations that are equal to the previous one do not represent a revision and are also excluded.

more is usually required to distinguish a normal distribution. The comparison is thus the core revision response to regular activity. I calculate the RRR by dividing the volumes after the revisions by the revisions before. This makes data comparable because it eliminates size effects. I express the above mathematically, for stock i at time t, as follows:

RRRi, t = !

ATVi, t : t+2 # ATVi, t-30 : t-1 ATVi, t-30 : t-1

$

I choose trading volumes over market returns for two reasons. Firstly, there is evidence that small investors buy more than they sell when a recommendation revision occurs, no matter the content of the new recommendation (Mikhail et al., 2007). This distorts the return data, because the confirmation bias should only induce investors to trade if the revision confirms their beliefs. If a positive response occurs at every revision, there is no difference between confirmation and disconfirmation. Trading volumes, on the contrary, show the magnitude of the response by combining the response of large and small investors. Secondly, trading volumes can be more informative about market responses and provide a relatively novel approach in many domains (Richardson, Tuna, & Wysocki, 2010). This type of data measures whether investors respond to the revision, rather than how they value it. Such a distinction is important for the confirmation bias because it induces investors to accept the information (if it is confirmatory of their beliefs) or disregard it. Market returns only capture the monetary value of the response of investors that accept the revision and respond to it.

The statistical tests comprise of comparisons of sample means of the RRRs as well as a simple regression analysis. For H1, RRRs to upgrades in a bull market are compared to

(23)

upgrades in a bear market. For H2, RRRs to downgrades in a bear market are compared to

downgrades in a bull market. Upgrades are not compared to downgrades because such an approach eliminates the factor of market sentiment. Also, prior research already documents disparities between responses to them (e.g., Mikhail et al., 2007). For the regression analysis, a simple dummy variable (BULL) suffices, because it is a supplemental test to the comparison of sample means. BULL takes a value of 1 in a bull market and 0 in a bear market. I do not include a variable for the absolute amount of the revision, because about 70% of the revisions are an upgrade or a downgrade of 1. For those observations, the variable would not explain differences among them. I compose the following regression equation for revision i at time t:

RRRi, t =

a

+

b

1 BULLt +

e

i

3.2 Anchoring bias

The effect of the anchoring bias is tested with a multiple regression analysis. First, I elaborate the sample. Since the hypotheses for this bias are not bound to a certain period, I examine observations within a period with market stability. This is favourable because the anchoring bias relates to a regular estimation process, which may be disturbed if there is market turbulence. In conformity with the bull market in the previous subsection, the data are from January 2012 up to December 2017. This time frame almost entirely consists of stable market growth (see appendix A). The firm selection is equal to that of the first two hypotheses, because the focus on large companies is similar. The analyses for the third and fourth hypothesis have the same limitations, except that the relevance of the firms is more constant. This is because I only examine one period, rather than several separate periods. Also, the time frame is recent, so most firms in the sample are likely to have been in the S&P 100 during the entire period.

The dependent variable is the investor’s estimation of a stock value. An adjustment to that estimate may lead to a trade, for instance when the analyst’s target price is increased further away from the actual stock price. If the anchor influences the value estimates of multiple investors and inclines them to trade, it becomes observable on the market. Therefore, target price publication responses (TPPRs) are measured with return data, adjusted for market returns.6_{Market return is proxied by the daily return on the S&P 500. I expect TPPRs to take}

considerable time to become observable, because a target price is an analyst’s expectation of

6_{This makes data more comparable because the core response is captured. For example, a 7% return that consists}

of 3% market return would equal a 7% return that contains 5% market return unless it is adjusted to 4% and 2% respectively.

(24)

the future value of the stock. Assuming that investors comprehend this, the response is represented by the cumulative abnormal return of stock i at time t (CARi, t), which is calculated

over a period of thirty trading days. Although target prices might cover a longer horizon, investors tend to respond to it within a shorter period. This period of thirty trading days starts on the day of the target price publication, so that the response on that day is included. Extension of this time frame may capture a continuance of the response, but it also increases the likelihood of interference of exogenous variables (for example, new target prices or other information). The return data for these periods are gathered from the CRSP database. Because this database restricts access to any data after 31 December 2017, I exclude all target prices that were published after 14 November 2017. This allows me to accumulate data of thirty trading days after the publication and leave one day out for data merging purposes.

The target prices are divided by the closing stock price on the day before the publication to obtain the “implied return” (Franck & Kerl, 2013, p. 2681). This describes the expected return by the analyst in the foreseeable future. The price of the day before the publication is used because the closing price on the day of the publication would include part of the market response to the target price. By dividing the target price by the actual price, any need for logarithmic scales dissipates, because an absolute variable becomes a relative one. The published target price, stock price and implied return of stock i at time t are denoted as

PTPi, t, pi, t-1 and IRi, t respectively, so:

IRi, t =

PTPi, t - pi, t-1 pi, t-1

Before I elaborate the remainder of the regression equation, I present the final sample size in Table 2. It is important to realise that the target price publications are not limited to revisions so they may include repetitions of prior values. In contrary to the recommendations, I do not filter revisions because the anchoring bias may affect investors even if the target price is equal to its predecessor. This is because an anchor is an implicit, unconscious effect so investors may be unaware of its presence. Therefore, it does not matter if the information is new or revised. Some observations with extreme implied or actual returns are removed from the sample. This is because the values are only extreme due to the occurrence of a stock split, a takeover or a similar distortion of price data. Also, since an IR requires a closing price on the day before, target prices published on the first trading day within the sample period are excluded. Similarly, target prices before an IPO cannot be converted to an IR and are therefore left out of the sample as well.

To supplement the anchoring effect and to partially control for exogenous factors, additional variables are added. The hypothesised effect of a PTP is that it triggers an

(25)

Table 2

Target price publications sample selection

Parts of sample Amounts

Initial amount of target price publications 50,890 Publications on the first day of the period or those before the IPO of the firm -986 Implied returns lower than -30% -143 Implied returns higher than 200% -36 Actual returns lower than -50% and higher than 150% -266

Final sample 49,459

Note. This table states the number of target price observations that are gathered from the I/B/E/S database. The initial amount of publications represents the raw data between 1January 2012 and 14 November 2017. Publications on the first trading day in the period or before an IPO are excluded because there is no closing stock price on the previous day. Implied and actual returns with extreme values are excluded because they relate to irregularcircumstances, such as stock splits and takeovers.

adjustment of the investor’s value estimate, which initiates at the PTP anchor. The adjustment process may include a comparison of the PTP with other outstanding target prices, or even a focus on them. Therefore, the average target price (ATPi, t), also known as the consensus or

the mean target price, close to the publication is included. The I/B/E/S database includes this consensus once a month, so the consensus date closest to the publication is used, rather than the most recent one, because it is most likely outdated.7_{Additionally, to control for dispersion}

between analysts’ target prices, their standard deviation denoted in dollars is included in the equation (ATP_STDi, t). This approach is similar to that of Bissessur and Veenman (2016) for

EPS forecasts. The ATP is divided by the closing price on the day before the new publication, so that its implied return (IR_ATPi, t) is obtained:

IR_ATPi, t =

ATPi, t - pi, t-1 p_{i, t-1}

The limitations of measuring an anchoring effect with market data are similar to those outlined in the introduction of this section. The two variables deducted above may simply measure correlation of the implied returns with actual returns. However, adding too many other variables that potentially support the anchoring effect may create an equation that is an unrealistic description of the process. The essence of the anchoring bias is, after all, that a single piece of information heavily distorts the value estimate. Hence, to strengthen the measurement of the bias to some extent, one variable is added to the equation. This variable is the outstanding average recommendation (ARi, t) which is obtained from the I/B/E/S

database, similar to ATP. It measures internal consistency of the analyst information, which may amplify the anchoring effect. For instance, if an investor observes a relatively high target

7_{For example, a consensus calculated 29 days ago is less accurate about the current situation than the consensus}

(26)

price, but many recommendations incline the investor to sell, the target price will unlikely lead to a stock purchase. Consequently, the anchoring effect is diminished. The recommendation scales are reversed, similarly to the previous subsection, so 1 is a strong sell and 5 is a strong buy.

To measure the precision or rounding effect of the PTP anchor, based on H4, another

variable is introduced. ROUNDING takes the same value as the number of zeroes at the end the target price. To illustrate, for a target price with no zeroes (e.g. $34.76) the variable equals 0. For a zero at the end (e.g. $34.40) it equals 1. For a target price that is rounded to the dollar (e.g. $34.00) it equals 2 and so on. Based on observations, most target prices are rounded to the dollar. However, further rounding is expected to impact the adjustment, so the variable is considered relevant. Also, I observe that a small selection of target prices are registered with three decimals. However, I do not incorporate this in the variable because stock prices are often rounded to cents as well.

I do not include any variables relating to size or industry, because the selection of firms is relatively small. Also, it comprises the hundred largest firms in the US so there is less size dispersion than if all companies would be included. Furthermore, no horizon variable is included. For every target price, I/B/E/S includes the amount of months that the analyst expects to pass by before the price reaches its target. The database denotes this as the horizon. However, this is often undisclosed, in which case a horizon of twelve months is assumed. This is clearly visible in the sample, because 99% of the target prices have that horizon. Therefore, the variable is not relevant. To conclude and summarise this subsection, the following regression equation is formed to test the anchoring bias:

CARi, t =

a

+

b

1 IRi, t +

b

2 IR_ATPi, t +

b

3 ATP_STDi, t +

b

4 ARi, t +

b

5 ROUNDINGi, t +

e

i

4 Results and analysis

In this subsection, I present the results of the tests. For the confirmation bias, I present descriptive statistics and regression output. Subsequently, I introduce frequency diagrams for

RRRs in each of the samples and elaborate them. For the anchoring bias, I first present

descriptive statistics and correlation coefficients. Second, I showcase the results of the regression analysis and discuss its implications. Third, I present an interval table that increases comprehensibility of the results.

(27)

4.1 Confirmation bias

The frequencies of the recommendation revisions are shown in Table 1 in the previous section. One interesting aspect that confirms market sentiment within the selected periods is the relative amount of upgrades and downgrades. In the bull market, upgrades make up the majority of the sample (53.2%). On the contrary, downgrades are the majority in the bear market samples (55.8%). These findings may simply relate to the presence of up- or downwards price and/or earnings movements, leading to revised expectations. It also signals slight analyst sentiment. However, Jegadeesh and Kim (2006) find that recommendation levels of analysts in the US are generally optimistic, consistently over time. The relative frequencies support their observations because analysts do not seem to completely adapt to different market circumstances. In other words, if analysts are indeed optimistic, the sample indicates that a bear market does not change that.

The descriptive statistics of the upgrade and the downgrade samples are presented in Table 3. It displays all individual sample means and standard deviations as well as the aggregates for both types of revisions. The first noticeable result is that all sample means are statistically different from zero. This is consistent with prior literature that documents significant responses to recommendation revisions (e.g., Chen & Cheng, 2006; Jegadeesh & Kim, 2006). The means represent percentage increases in ATVs. For instance, if a stock recommendation is upgraded, trading volumes increase by 29.40% on average during the three-day period after the revision in comparison to the thirty-day trading period before. In all cases, trading activity increases. However, the percentages are lower than those found by Mikhail et al. (2007, p. 1238). Their tests showcase 70% and 100% increases for large and small investors respectively, which increases further as the absolute value of the revision itself increases. This difference in findings may be explained by the fact that my sample focuses on the 100 largest firms in the US, while their sample is unlimited. Larger firms are presumably more stable and therefore their stock is less volatile than, for example, start-ups. This in turn may lead to investors of larger firms being less sensitive to recommendation revisions. Also, larger firms are often covered by more analysts, so that there is a higher probability that the information value of a single revision is diluted. In conclusion, the descriptive statistics confirm prior findings that investors respond to recommendation revisions, albeit weaker than in full market samples.

Table 3 also states the differences between the sample means. The means do not significantly differ from each other, based on independent samples tests. The differences also represent the coefficients of the BULL variable in both simple regression analyses, which are insignificant as well. All regression output is therefore presented in Appendix B. Based on these results, I do not find direct statistical support for H1 and H2. However, the signs of the

Investors’ cognitive bias and sell-side analyst information

1