Dissecting Momentam

Dissecting Momentum by Luís Carlos Pereira Magalhães Dissertation of Masters in Finance and Taxs Advisers: Dr. António Cerqueira Dr. Elísio Brandão 2013 DISSECTING MOMENTUM Author: Luís Magalhães (a) School of Economics and Management, University of Porto, Portugal Advisers: António Cerqueira (b) School of Economics and Management, University of Porto, Portugal Elísio Brandão (c) School of Economics and Management, University of Porto, Portugal E-mail: (a) [email protected]; (b) [email protected]; (c) ebran[email protected]t September 2013 i ABSTRACT Several studies provide evidence that the effect of most anomalies tend to disappear as soon as they are discovered, but that was not the case of Momentum. After two decades of research, Momentum still exists on stock markets and with the same strength evidenced in Jegadeesh and Titman (1993). The goal of this paper is to bring a new perspective of Momentum spectrum. Our finding that many stocks change from “Winner” to “Loser” frequently seems to support the hypothesis that firm specific characteristics are not expected to be the key determinant of Momentum classification. However, the main result of this study is that the understanding of the sources of Momentum anomaly requires taking into account two kinds of stocks: those who are frequently “Players" (“Winner” or “Loser”) and those that rarely are. Our study provides empirical evidence on a number of factors that explain the classification of firms as “Players” or “Non Players” based on Discriminant Analysis and Logit Regression. At our knowledge, we are the first to provide such kind of evidence. Keywords: Momentum Classification, Firms’ Characteristics JEL Codes: G11, G12, G14 ii Index 1. Introduction ................................................................................................ 1 2. Literature Review ....................................................................................... 4 3. Data and Methodology ............................................................................... 6 4. Results ........................................................................................................ 7 4.1. Momentum Profitability .......................................................................... 7 4.2. Momentum Groups: PG and NPG ......................................................... 15 4.3. Discriminant Analysis ........................................................................... 21 4.4. Sorting Analysis .................................................................................... 23 4.5. Logit Regression ................................................................................... 26 5. Conclusions .............................................................................................. 30 6. Variables Description ............................................................................... 31 7. References ................................................................................................ 33 Appendix ............................................................................................................. 35 iii Panel Index A. Descriptive Statistic of Momentum ............................................................ 8 B. Average Returns by Civil Year Months ..................................................... 9 C. Crisis Effect and Momentum ...................................................................... 6 D. Descriptive Statistic after Fama and French Adjustment ......................... 12 E. After Adjustment Average Returns by Civil Year Months ........................ 7 F. Crisis Effect after Fama and French Adjustment ...................................... 7 G. Stocks in Momentum Classification ......................................................... 16 H. Adjustment changes on “Winners” and “Losers” ................................... 17 I. Correlation Analysis ................................................................................. 17 J. Discriminant Analysis .............................................................................. 22 K. Sorting Approach ...................................................................................... 24 L. Logit Regression ....................................................................................... 27 1 1. Introduction Since Jegadeesh and Titman (1993) seminal paper, Momentum anomaly has been a subject of many papers in Financial Literature. More specifically, two different (but related) trends appeared: one trying to quantify its profitability and another focused on finding its sources. The first one basically shows that the Strategy of buying past “Winners” (stocks with the best performances over past 6 months) and selling “Losers” (those with the worst performance over past 6 months) is profitable in many different markets all over the world (see, for example, Rouwenhorst (1998, 1999) for Europe and Emerging Markets and Chui, Titman and Wei (2003) for Asian Markets) and for different time periods (e.g. Jegadeesh and Titman (2001) and Fama and French (2008, 2012)). On the other side, several works provide different explanations to the source of this anomaly. Firms’ characteristics like Size, Book-to-Market ratio (see, for example, Hong, Lim and Stein (2000) and Fama and French (2008, 2012)1), and Credit Risk (e.g. Avramov, Chordia, Jostova and Philipov (2007)) to risk based explanations (see, for example, Fama and French (1996) and Griffin, Ji and Martin (2003)), arbitrage limits (e.g. Shleifler and Vishny (1997)) and more possible causes were appointed to it. However, there is not a consensus about the source(s) of the anomaly and we think we could bring a new perspective of Momentum spectrum. We extend Momentum Literature in three ways by showing that: (1) contrarily of we could think, it seems that firms specific characteristics are not the key to the classification of stocks, (2) there is a positive and significant correlation between the number of times that a company is classified as “Winner” and “Loser” and (3) there are two kinds of companies: those who are frequently “Players” (stocks classified as “Winner” or “Loser”) and those where the likelihood of being classified as one of them is very low. At our knowledge, we are the first to provide such kind of evidence. According to our results, almost all stocks were classified at least once as “Winner” (93.30% using non adjusted returns and 84.60% with abnormal returns2) or “Loser” (78.89% and 81.97%, respectively) for the sample period from January 2003 to 1 Their results shows that Momentum profitability is connected to small caps but this fact cannot explain its existence 2 Resids from Fama and French Three Factor Model Adjustment (Fama and French (1993)) 2 December 2011. This finding, connected to the discovery of concentration of “Players” observations in a small group of stocks, seems to suggest that firm specific characteristics cannot explain, by themselves, the classification of Momentum Strategy. Initially, we thought about ranking firms by the number of times that a stock was classified in each kind in order to dissociate Top “Winners” from Top “Losers” (Top quintile of each rank). However, that becomes impossible because some stocks were in both, so we had no other option than studying Top “Players” (Top quintile of “Players” ranking). With a selection criterion, we thought it would be possible to separate them in three groups: Winners, Losers and The Others. However, we did not find one capable of have balanced groups (in this context, balanced means with, approximately, the same number of stocks in each one) and, at the same time, make some sense in the dissociation so, our results appointed, clearly, to the impossibility of dissociation. Just a few stocks are, clearly, “Winners” or “Losers” because the majority of them change often in their classification. In order to finish this issue, we verified the correlation between the number of times that a company is classified as “Winner” and “Loser” and, contrarily to our initial expectations (but expected after verify the impossibility of dissociation), this correlation is positive (0.4387) and very significant (t = 16.21). This shows, without doubts, that we cannot dissociate these two kinds of “Players”. However, how can Momentum be so profitable when we are not capable to dissociate the two kinds of “Players”? Why firms’ characteristics (such as Size (Fama and French (2008, 2012)) and Credit Risk (Avramov, Chordia, Jostova and Philipov (2007)) have impact in the profitability of Momentum Strategy? The answer to these questions can be the existence of two groups of “Players”. When we rank the sample by number of times that a stock is “Player”, the bottom quintile represents only 1.84% of total “Players’” observations and the top quintile has 44.54% of it. As we can see, the difference to the percentage of stocks in each quintile (i.e. around 20%) is very high so, according to our results, we can conclude that there are two different groups: The “Players” (PG and top quintile) and The “Non-Players” (NPG and bottom quintile). In this way, firms’ characteristics as Size and Credit Risk cannot explain Momentum but can be capable to explain the existence of these two groups. 3 In order to show if this information is correct, we saw if those characteristics are capable to dissociate the two groups in a Discriminant Analysis. The most important results of this approach are that Size, Credit Rating and firm’s Profitability are the key to dissociate PG from NPG. Spread can be efficient on dissociation because of its coefficient but, as the values that it assumes are very low, this is not a key characteristic. On the other hand, Book-to-Market, even being statistically significant, does not have discriminatory capacity. All firms’ characteristics analyzed have a positive sign except for Size. To finalize our paper, we present a Logit regression to get the probability of a stock being classified as “Player” in the period. The results of this approach are similar (in sign and importance) to Discriminant Analysis, with the exception of Spread that has a negative coefficient. Nevertheless, this variable has not a significant impact in the likelihood of stocks’ classification because, in our opinion, of the information contained on Dummy variables3. We present also a Sorting analysis to get the marginal effects in the likelihood of a stock be a “Player. However, Logit Regression cannot predict if a stock will be a “Player” in the next period because, to have a likelihood of at least 50%, it is necessary a small cap with negative net income. By the other side, Sorting analysis suggests that it is not necessarily the stock be a small cap because the coefficient of the Size bottom quintile is not statically significant. So, we can easy conclude that other sources not related with the characteristics of “Player” may be capable of get better results in this domain. The remainder of this paper is organized as follow: Section I brief reviews Momentum Literature, Section II presents the data and methodology of this paper. Section III contains the results in the following order: first of Discriminant Analyzed followed by a Sorting approach and finalize with a Logit Regression. Finally, Section IV concludes. 3 Spread is a proxy for information assymetry and transaction costs. As transaction costs do not have influence in Momentum (see Korajczyk and Sadka (2004) for more information), the impact of this variable just can be connected to information. 4 2. Literature Review Since Jegadeesah and Titman (1996) seminal paper, Financial Literature tries to discover the sources of Momentum anomaly. However, at this moment, there are not many consensuses about this. As an anomaly disappears of the markets with the help of arbitrageurs, the risk factor is one of the most studied as source of this anomaly because, if there is some kind of risk in Momentum Strategy, they will not act and the profitability of this trading strategy can be explained by this. However all papers that studied this question concludes the same: Risk do not have influence in Momentum. Indeed, Fama and French (1996) shows that this anomaly is the only that their Three Factor Model cannot explain. In the same line, Shleifler and Vishny (1997) study if arbitrage limits can explain the existence of Momentum. This finding is very important to this question for the same reason of risk: if there are limits to arbitrage, arbitrageurs will not be capable to act and lead to the end of this anomaly. However, they argue that the limits exist but, in the limit, are capable to help to explain the persistence of Momentum. As arbitrageurs can act, other sources were appointed to this anomaly. The impact of information is also a trend very studied in the Literature because of its uncertainty. Zhang (2006) is one the most complete works in this domain by showing that the profitability of Momentum has a positive correlation with information uncertainty and, more specifically, is related to an overreaction to new information when its uncertainty is higher. This finding is crucial because prior Literature attributed price continuation to a slow market response to information (see, for example, Chan, Jegadeesh and Lakonishok (1996)) and he shows exactly the opposite. Another important finding is the effect of Size. Several works (including Fama and French (2008, 2012)) show that the profitability of Momentum is related to small caps. Of all firms characteristics appointed to Momentum “Players” this is the only where exist some consensus. As for the Credit Risk effect, we think that Avramov (2007) work is very important. He argues that stocks with a high grade of credit rating do not exhibit Momentum contrarily of the low grade stocks. In fact, he goes further: this specification is independent of Size because even small caps with high grade do not exhibit Momentum and big stocks with 11 Looking more profoundly to this question, we can see that it is related to the performance of “Winners” that decreased 11.55% (1.93% per month and t = -2.51). On the other hand, the performance of “Losers” seems to be more stable even in recession periods because it only increased 2.28% (0.38% per month and t = 0.38). A possible explanation to this can be a change in specifications of “Players”. To eliminate risk factor, we have adjusted monthly returns for each company following Fama and French Three Factor Model (Fama and French 1993) and results show that risk may explain Momentum anomaly (Panel D). However, this result is not in coherence with previous Literature (see, for example, Fama and French (1996) and Griffin, Ji and Martin (2003)). The reason to this fact is the actual financial crisis because, as we can see in Panel F, this strategy was profitable at the beginning of the year 2008 in EW (1.43% and t = 7.81) and VW (1.70% and t = 4.37). So, we could conclude that Momentum profitability can be related to firms’ characteristics (because their changes in recession periods) but it has also a relation with macroeconomics effects. Nevertheless, we believe that if we analyze a larger period, the impact of this recession will be lower and, consequently, Momentum Strategy will have positive returns after Fama and French Adjustment, excluding, this way, risk as a possible cause. However, we must remember that Griffin, Ji and Martin (2003) argue that we can exclude macroeconomics effects from Momentum explanations. To finalize Momentum profitability analysis, we want to pay a little attention to Panel E. As we can see, this trading strategy, even after Fama and French Three Factor Model Adjustment, just produces statically significant returns after accounting reports. However, this fact is not as clear as before adjustment but still exists. This brings some robustness to hypotheses of “Players” being stocks where these reports are the almost the only source of information that common investors can get. On other hand, we cannot forget that these results are not very robust because of the low number of observations. 12 Panel D - Descriptive Statistic after Fama and French Adjustment Portfolio µ Max Min Med Σ Equal-Weight (EW) W -4.23% (-2.92)*** 20.30% -57.76% -2.56% 0.1420 L -6.08% (-3.76)*** 62.34% -23.86% -10.25% 0.1585 (W-L) 1.85% (1.10) 25.87% -57.07% 6.82% 0.1651 (W-L)/6 0.31% (1.10) 4.31% -9.51% 1.14% 0.0276 Value-Weight (VW) W -5.19% (2.44)** 31.87% -103.82% -2.19% 0.2083 L -8.66% (-3.98)*** 70.74% -45.02% -9.31% 0.2132 (W-L) 3.47% (1.40) 46.79% -76.65% 4.85% 0.2426 (W-L)/6 0.58% (1.40) 7.80% -12.78% 0.81% 0.0404 Similar to Panel A, we present some descriptive statistics about Momentum portfolios. The difference to Panel A is that this one shows the results after Fama and French a Adjustment (Fama and French (1993)). In brackets we have t-statistics and its significance. *, ** and *** represents 10%, 5% and 1% of significance, respectively. 13 Panel E - After Adjustment Average Returns by Civil Year Months Portfolio Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec Equal-Weight or EW W -8.57% (-1.86) -11.13% (-1.93)* -10.27% (2.16)* -14.26% (2.65)** -12.08% (-1.65) -11.05% (-1.36) -1.77% (-0.26) -1.37% (0.36) -1.12% (-0.33) -0.28% (-0.09) 0.86% (0.30) -1.36% (0.52) L -3.86% (-0.50) -7.26% (-0.94) -5.73% (-0.87) -9.24% (-1.77) -12.20% (-2.94)** -15.14% (-4.19)*** -6.69% (-1.85) -8.16% (-2.66)** -9.45% (-2.53)* -1.78% (-0.22) -1.81% (-0.21) 1.07% (0.10) (W-L) -4.71% (-0.73) -3.87% (-0.71) -4.54% (-0.84) -5.02% (-0.82) 0.12% (0.02) 4.09% (4.08)*** 4.91% (1.09) 6.79% (3.88)*** 8.34% (1.88)* 1.50% (0.16) 2.67% (0.32) -2.43% (-0.26) (W-L)/6 -0.78% (-0.73) -0.64% (-0.71) -0.76% (-0.84) -0.84% (-0.82) 0.02% (0.02) 0.68% (4.08)*** 0.82% (1.09) 1.13% (3.88)*** 1.39% (1.88)* 0.25% (0.16) 0.44% (0.32) -0.40% (-0.26) Value-Weight or VW W -2.52% (-0.57) -6.01% (-1.79) -10.24% (-1.59) -10.73% (-1.46) -13.57% (-0.92) -12.65% (-0.97) -10.08% (-0.86) -1.57% (-0.38) 2.86% (0.53) 7.43% (1.18) 1.80% (0.44) -7.04% (-2.66)** L -10.55% (-1.08) -5.40% (-0.47) -5.66% (-0.61) -15.50% (-2.56)** -10.47% (-1.75) -4.70% (-0.66) -8.10% (-1.98)* -12.84% (-1.88)* -13.77% (-2.82)** -6.39% (-1.09) 0.92% (0.11) -11.49% (-1.11) (W-L) 8.03% (0.78) -0.61% (-0.06) -4.58% (-0.73) 4.77% (0.81) -3.10% (-0.33) -7.95% (-0.72) -1.98% (0.23) 11.27% (2.61)** 16.64% (2.83)** 13.82% (1.68) 0.88% (0.09) 4.45% (0.39) (W-L)/6 1.34% (0.78) -0.10% (-0.06) -0.76% (-0.73) 0.80% (0.81) -0.52% (-0.33) -1.32% (-0.72) -0.33% (-0.23) 1.88% (2.61)** 2.77% (2.83) 2.30% (1.68) 0.15% (0.09) 0.74% (0.39) Similar to Panel B, this panel shows the returns of 6 Months Momentum for each Civil Year Months for both approaches (EW and VW). The difference to the other Panel is that in this one, we considered the abnormal returns of Fama and French (1993) Adjustment. In brackets we have t-statistics and its significance. *, ** and *** represents 10%, 5% and 1% of significance, respectively. 14 Panel F - Crisis Effect after Fama and French Adjustment Portfolio Before December 2007 (1) After January 2008 (2) Difference (2) - (1) Equal-Weight or EW W -2.67% (-2.20)** -6.24% (-2.14)** -3.57% (-1.13) L -11.26% (-11.91)*** 0.57% (0.18) 11.82% (3.51)*** (W-L) 8.59% (7.81)*** -6.81% (2.18)** -15.39% (-4.64)*** (W-L)/6 1.43% (7.81%) -1.13% (2.18)** -2.57% (-4.64)*** Value-Weight or VW W -1.94% (-1.18) -9.38% (-2.17)** -7.44% (-1.61) L -12.12% (-6.50)*** -4.22% (-0.98) 7.90% (1.69)* (W-L) 10.18% (4.37)*** -5.16% (-1.15) -15.34% (-3.03)*** (W-L)/6 1.70% (4.37)*** -0.86% (-1.15) -2.56% (-3.03)*** Similar to Panel C, this panel shows the returns of 6 Months Momentum before and after 2008 finacial crisis for both approaches (EW and VW). The difference to the other Panel is that in this one, we considered the abnormal returns of Fama and French (1993) Adjustment. In brackets we have t-statistics and its significance. *, ** and *** represents 10%, 5% and 1% of significance, respectively. 15 4.2. Momentum Groups: PG and NPG However, our results seem to suggest that firms’ specific characteristics do not have influence in the classification of Momentum “Players”. When we analyze the percentage of stocks that were classified as “Winners” or “Losers” along the sample period, we will see that this percentage is very high in both kinds of “Players” in all approaches (see panel G). This finding connected with information of Panel I (concentration of “Players” observations in a small group of stocks) suggests that firms’ specific characteristics do not have a lot of influence in the classification of stocks. As the stocks used in this anomaly are not the same with and without Fama and French Three Factor Model Adjustment, it is important to prove that they do not differ a lot from each other and we can find this information on Panel H. Only 7.8% of total observations changed their classification and around 20% of “Players” were “declassified”. This way, we can affirm that they not differ significantly from each other so, by now on, we will just analyze after adjustment results. With this information, we thought that we could be capable to dissociate Top “Winners” and Top “Losers” from the others but, to our surprise, several stocks were in both. So, we decided to separate Top “Players” in three groups (Winner, Loser and Others Groups) with a selection criterion. However, we could not find one capable of get balanced groups (in this context, balanced means with, approximately, the same number of stocks in each one) and, at the same time, capable of make some sense. The reason to this fact is that the number of time that each stock is classified in one kind of “Players” is close to the number of times classified in the other. To confirm this, we calculated the correlation between a numbers of times that a stock was considered “Winner” and “Loser”. As we were expecting now, the correlation is positive (0.4387) and very significant (t = 16.21) confirming symmetry hypotheses12. At our knowledge, we are the first to provide this kind of evidence. In order to have more robustness, we present also, in Panel I, a “correlation analysis”. Basically, we try to get the spectrum of the correlation in each quintile and we join Size to have a foreign characteristic13 in this analysis. 12 As the number of times that a stock was considered in a kind increases, the number of times that it was in the other increases either. 13 In this context, foreign means not related with “Players” classification 16 Panel G - Stocks in Momentum Classification Group Ηi ηt ηi/ηt Before Fama and French Adjustment Equal-Weight or EW W 1030 1104 93.21% L 871 1104 78.82% P 1058 1104 95.75% Value-Weight or VW W 1019 1104 92.30% L 862 1104 78.08% P 1053 1104 95.38% After Fama and French Adjustment Equal-Weight or EW W 934 1104 84.60% L 905 1104 81.97% P 1009 1104 91.39% Value-Weight or VW W 924 1104 83.70% L 892 1104 80.80% P 1004 1104 90.94% This panel shows information about the number and percentage of stocks that had been a part of Momentum Classification before and after Fama and French Adjustmente (Fama and French (1993). ηi and ηt represents the number of different stocks classified in each group and the total of stocks, respectivelly. W, L and P represents “Winners”, “Losers” and “Players” 17 Panel H - Adjustment changes on “Winners” and “Losers” Portfolio Changes Total Percentage Total or T W 4413 103835 4.25% L 3686 103835 3.55% Partial or P W 2230 9453 23.59% L 1842 9411 19.57% This panel shows the number (and its percentage) of companies that changed their classificafion because of Fama and French Adjustment (Fama and French (1993)). In this context, Total (or T) means the total changes in classification of all stocks and Partial (or P) means the changes in original “Players”. W and L means “Winners” and “Losers”, respectivelly Panel I: Correlation Analysis Q η %%η %%P %%(P-η) Correl (W,L) Winners Sort 1 239 21.65% 3.56% -18.09% 0.1305 (2.03)** 2 239 21.65% 12.51% -9.13% 0.2743 (4.39)*** 3 193 17.48% 16.84% -0.65% 0.0750 (1.04) 4 226 20.47% 29.82% 9.35% 0.0983 (1.48) 5 207 18.75% 37.28 18.53% -0.0209 (-0.30) Total 1104 100% 100% 0% 0.4387 (16.21)*** 18 Q η %%η %%P %%(P-η) Correl (W,L) Losers Sort 1 262 23.73% 4.70% -19.04% 0.1931 (3.17)*** 2 223 20.20% 12.11% -8.09% 0.1871 (2.83)*** 3 201 18.21% 17.77% -0.43% 0.0801 (1.13) 4 198 17.93% 24.86% 6.93% 0.0465 (0.65) 5 220 19.93% 40.56% 20.63% 0.0204 (0.30) Total 1104 100% 100% 0% 0.4387 (16.21)*** Players Sort 1 246 22.28% 1.98% -20.30% -0.1657 (-2.62)*** 2 211 19.11% 8.93% -10.18% -0.7444 (-16.12)*** 3 218 19.75% 18.09% -1.66% -0.8132 (-20.54)*** 4 214 19.38% 27.77% 8.39% -0.8721 (-25.94)*** 5 215 19.47% 43.23% 23.75% -0.6013 (-10.98)*** Total 1104 100% 100% 0% 0.4387 (16.21)*** 19 Q η %%η %%P %%(P-η) Correl (W,L) Size Sort Small Caps 221 20.02% 29.49% 9.47% 0.2538 (3.88)*** 2 221 20.02% 25.62% 5.60% 0.3273 (5.13)*** 3 220 19.92% 21.29% 1.37% 0.4249 (6.93)*** 4 221 20.02% 16.23% -3.79% 0.3512 (5.55)*** Big Stocks 221 20.02% 7.36% -12.66% 0.4019 (6.49)*** Total 1104 100% 100% 0% 0.4387 (16.21)*** This panel shows information about the correlation between the number of times that a stock is classified as “Winner” and “Loser” in four sorts (Winner, Loser, Player and Size). In the first collumn we have the quintiles (Q). In the following collumns we have, respectivelly, information about the number of stocks (η), its percentage (%%η), the percentage of total observations of “Players” observations (%%P), the difference between the last two collumns (%%(P-η)) and the correlation factor of the number times that a stock is classified in each kind of “Players” (Correl(W,L)) for each quintile. In brackets we have t-statistics and its significance. *, ** and *** represents 10%, 5% and 1% of significance, respectively. 20 As expected, there is a negative correlation when we sort in “Players”. In fact, no other result was expected because this category is the sum of times that a stock is classified as “Winner” and “Loser” so, if we sort it by “Players”, the only possible result is a negative correlation between the number of times that it was “Winner” and “Loser”. However, when we sort in one of the kinds of “Players”, the conclusion is not the same. As Panel I shows, the correlation is always positive14 but just in the lower quintiles is significant. On the other hand, in both sorting, there is a positive effect of increasing the classifications as “Winner”/“Loser” in the other kind exposing, this way, the relation between them. This fact is totally confirmed by the sorting on Size. As we can see in same Panel, the correlations between them are positive and very significant in all quintiles of this characteristic. So we can conclude without doubts that exists a positive relation between these two phenomena. Another inference, as we explained above, is that firms’ specific characteristics do not seem to be the key of the classification of Momentum “Players”. But, if that is true, why do characteristics as Size and Credit Rating have some impact in Momentum Strategy? The answer can be on the existence of two groups of stocks: The Player (or PG) and The Non Players (or NPG). The first one is composed by stocks that are frequently classified as “Players” and the other by stocks were the likelihood of being classified as “Player” is very low. We can get this information from Panel I, more specifically from “Players” Sorting. The bottom quintile only represents 1.98% of total “Players” observations but the top quintile represents 43.23%. As the difference to the percentage of stocks is very high in booth quintiles, we can easily conclude that there exist two groups: PG and NPG. At our knowledge, we are the first to provide such kind of evidence. This discovery could be crucial to understand Momentum spectrum. In the limit, we were looking to the wrong perspective of it and, perhaps, now, with this new information, we may find the reason to the existence of this anomaly. However, before that, future studies need to confirm these findings around the world and along the time in order to eliminate the remote possibility of this fact be specific of London Stock Exchange or of the our sample period. 14 The only exception is the top quintile when sorted by number of times classified “Winner” but is not statiscally significant 27 Panel L: Logit Regression Variable Coeficient C 0.359112 (4.97)*** Size (t-6) -0.231682 (-39.38)*** BTM (t-6) 0.014669 (3.63)*** Spd (t-6) -0.401822 (-2.31)** HG (t-6) 0.407045 (9.10)*** LG (t-6) 0.320877 (6.33)*** NegNI (t-6) 0.951130 (30.69)*** HP (t-6) 0.339521 (7.29)*** LP (t-6) 0.271915 (6.38)*** HG (t-6) x NegNI (t-6) -0.097107 (-1.77)* HG (t-6) x HP (t-6) 0.107526 (1.19) HG (t-6) x LP (t-6) -0.011064 (-0.14) LG (t-6) x NegNI (t-6) -0.044525 (-0.75) LG (t-6) x LP (t-6) 0.089835 (1.02) LG (t-6) x HP (t-6) -0.100113 (-1.07) Observations 85186 LR Statistic 7828.714*** McFadden R² 0.095381 Log likelihood -37124.94 This panel shows the results of the final Logit Regression. In brackets we have z-statistics and its significance. *, ** and *** represents 10%, 5% and 1% of significance, respectively. 28 On the other hand, Spread and BTM (because of the values that both, generally assume) do not seem to have any influence in Momentum classification but, if this was expected to BTM, Spread should have some significance and other sign. This variable, by itself, has discriminative (Discriminant Analysis) and explanatory (Sorting Approach) capacity. The real question is: why does this variable fail in the other models? In our opinion, this may be because of information. As we know, Spread is a proxy for transactions costs and information asymmetry and it just could be important by information because transaction costs are insignificants to this question (see, for example, Korajczyk and Sadka (2004)). We understand that Dummies are information to the market and the truth is that they are all statistically significant. In our opinion, Spread’s insignificance is related to these variables: they get the main effect of lack of information. Confirm this hypothesis; in the model without dummies for information, the sign of this variable coefficient are positive (see Appendix B)17. This result is, in a certain way, empirical support to the hypothesis of information as key of Momentum classification. However, our model has a low capacity to obtain the Good News effect (in this context, we consider good new HG and HP). That is because it is not very easy to know what is or is not a good new. See the example of Apple: their accounting report shows an historic positive net income in 2012 and theirs stocks had strong negative returns because of that. What investors understand as a good new is more complicated to put in a regression than a bad new (difficultly, negative net income (NegNI) and low credit rating (LG), for example, are not understand as a bad new). In further works, it is crucial to have a better methodology to get this kind of information and produce a model were it is possible to predict which stocks will be “Players” in the period. Another important thing is to have more and better variables capable of measuring information because, once again, our results appoint to that as a possible explanation of Momentum Classification. In this regression, we analyze also the cumulative effects of Information (i.e. the coexistence of two sources of information). The interaction between Dummy variables will tell us if the coexistence of news has any kind of effect in the likelihood of a stock be considered “Player” in the period and if it is positive or negative. 17 Its coefficient is not statistically significant is this apporach. However, in all other, we can conclude that Spread has some impact in Momentum Classification 29 As we can see in Panel L, interactions between Dummies are not very significant to the likelihood of being a Momentum “Player” (just the interaction between NegNI and HG is significant but this result is not very robust). However, with a short period as ours, this can be normal because, to have an interaction, it is need that a stock is, at the same time, in top and/or bottom quintiles of two different ranking. If we consider that every stock has the same probability of being there, just around 10%18 of total observations have conditions to have multi-effects and they will be divided in six variables. With such a short number of observations it is very unlikely that these variables have significance in a Logit Regression. We believe that, with a better way of getting information, all these variables will be significant so it is more important to see the sign of them than its significance. The most important result of these interaction is that there positive cumulative effects when both Dummies represents the same kind of news and that a good new combined with a bad new has a negative cumulative effect. However, these results were expected because it is normal that a good (bad) new combined with another one produce a higher effect than when they are isolated and, it is even more acceptable, that a good combined with a bad leads to a lower effect because they will eliminate, in a certain way, the effect of one of them. Here it would be interesting to know what effect and in what conditions that effect remains in the market. In sum up, our model has explanatory capacity but cannot predict witch stocks will be classified as “Players” because it has difficulties in measuring the effects of good news and size effect is exaggerated (Sorting approach shows that this variables is significant in the Top quintile and not in the Bottom). Another problem is the low number of observations in the interaction variables so it is needed to improve this model to confirm our results. Finally, these results appoints to the hypothesis of information as key of Momentum Classification. 18 We have three dummies to Profitability and one of them (NegNI) has around 30% of total observations 30 5. Conclusions In our opinion, the main goal of this paper is accomplished: we bring a new perspective of Momentum spectrum when we show that there are two different groups of Stocks (PG and NPG). We argue that Size as the main characteristic to distinguish “Players” from the others (and, consequently, PG from NPG) but cannot do it by itself because information (measured by dummy variables) pays a roll of importance in this domain. We confirm also that Book-to-Market is insignificant after Fama and French Adjustment (Fama and French (1993)) and Spread effect is limited because it is only important do not have a low relative Spread (but not necessarily a big one) to have a higher likelihood of being classified in the period. Further works may try to eliminate our limitations. If some of them are easily solved (as reply this work in other market and in different (and larger) periods of time) others will represent a better challenge. Test the hypothesis of information as key of Momentum classification will be one of the most challenging of all because it is very difficult to measure information. It could be also interesting evaluate the discriminatory capacity of other variables as, for example, analyst coverage and others firms characteristics. Nevertheless, our work is very important to Momentum Literature because it can lead to the answer that everyone tries to get: Why Momentum exists? We cannot forget that, even with those limitations, our findings are robust because all approaches lead to the same conclusion and, for the first time, we know that there are two different groups in the Markets: PG and NPG. 31 6. Variables Description Size: Logarithm of Market Capitalization Size = ln 𝑀𝐶 where MC means Market Capitalization (in Millions) BTM: Book-to-Market ratio 𝐵𝑇𝑀= 𝐵𝑉 𝑀𝐶 where BV means Book-Value and MC Market Capitalization (both in Millions) Spd: Relative Spread 𝑆𝑝𝑑 = Spread 𝑃 where Spread is half of the difference between the Ask and Bid Price and P is stocks Price HG: High Grade Stocks 𝐻𝐺 = 1, 𝑖𝑓 𝑖𝑠 𝑖𝑛 𝐶𝑅 𝑇𝑜𝑝 𝑄𝑢𝑖𝑛𝑡𝑖𝑙𝑒 0, 𝑜𝑡𝑕𝑒𝑟𝑤𝑖𝑠𝑒 where CR means Credit Risk. CR is measured by BSM model (Black and Scholes (1973) and Merton (1974)) 32 LG: Low Grade Stocks 𝐿𝐺 = 1, 𝑖𝑓 𝑖𝑠 𝑖𝑛 𝐶𝑅 𝐵𝑜𝑡𝑡𝑜𝑚 𝑄𝑢𝑖𝑛𝑡𝑖𝑙𝑒 0, 𝑜𝑡𝑕𝑒𝑟𝑤𝑖𝑠𝑒 where CR means Credit Risk. CR is measured by BSM model (Black and Scholes (1973) and Merton (1974)) NegNI: Stocks with Negative Net Income 𝑁𝑒𝑔𝑁𝐼 = 1, 𝑖𝑓 𝑁𝐼 < 0 0, 𝑜𝑡𝑕𝑒𝑟𝑤𝑖𝑠𝑒 where NI means Net Income HP: High Profitability 𝐻𝑃 = 1, 𝑖𝑓 𝑖𝑠 𝑖𝑛 𝑅𝑂𝐸 𝑇𝑜𝑝 𝑄𝑢𝑖𝑛𝑡𝑖𝑙𝑒 0, 𝑜𝑡𝑕𝑒𝑟𝑤𝑖𝑠𝑒 where ROE means Return on Equity (Net Income/Equity). Stocks with Negative Net Income were excluded because they are considered in NegNI variable LP: High Profitability 𝐿𝑃 = 1, 𝑖𝑓 𝑖𝑠 𝑖𝑛 𝑅𝑂𝐸 𝑇𝑜𝑝 𝑄𝑢𝑖𝑛𝑡𝑖𝑙𝑒 0, 𝑜𝑡𝑕𝑒𝑟𝑤𝑖𝑠𝑒 where ROE means Return on Equity (Net Income/Equity). Stocks with Negative Net Income were excluded because they are considered in NegNI variable 33 7. References Avramov, Doron, Tarun Chordia, Gergana Jostova, and Alexander Philipov, 2007, “Momentum and Credit Risck”, Journal of Finance, 62, 2503-2520 Black, F. and M. Scholes, 1973, “The Pricing of Options and Corporate Liabilities”, Journal of Political Economy, 81 (1973), 637-59 Cooper, Michael J., Roberto C. Gutierrez Jr., and Allaudeen Hameed, 2004, “Market States and Momemtum”, Journal of Finance, 59, 1345-1365 Chan, Louis K.C., Narasimhan Jegadeesh, and Josef Lakonishok, 1996, Momentum strategies, Journal of Finance, 51, 1681-1713 Chui, C.W. Andy, Sheridan Titman, and K.C. John Wei, 2003, “Momentum, legal systems and ownership structure: An analysis of Asian stock markets”, Working paper, University of Texas at Austin. Chui, Andy C. W., Sheridan Titman and K.C. John Wei, 2010, “Individualism and Momentum around the World”, Journal of Finance, 65, 361-392 Daniel, Kent, David Hirshleifer, and Avanidhar Subrahmanyam, 1998, “Investor psychology and security market under- and overreactions”, Journal of Finance 53, 1839–1886 Fama, Eugene F., and Kenneth R. French, 1993, “Common risk factors in returns on Stocks and Bonds”, Journal of Financial Economics, 33, 3-56 Fama, Eugene F., and Kenneth R. French, 1996, “Multifactor Explanations of Asset Pricing Anomalies”, Journal of Finance, 51, 55-84 Fama, Eugene F., and Kenneth R. French, 2008, “Dissecting Anomalies”, Journal of Finance, 63, 1653-1678 34 Fama, Eugene F., and Kenneth R. French, 2012, “Size, value and momentum in international stocks returns”, Journal of Financial Economics, 105, 457-472 Griffin, John M., Xiuqing Ji, and J. Spencer Martin, 2003, “Momentum Investing and Business Cycle Risk: Evidence from Pole to Pole”, Journal of Finance, 58, 2515-2547 Hong, Harrison, Terence Lim and Stein C. Jeremy 2000, “Bad News Travels Slowly: Size, Analyst Coverage, and the Profitability of Momentum Strategies”, Journal of Finance, 55, 265-295 Jegadeesh, Narasimhan, and Sheridan Titman, 1993, “Returns to Buying Winners and Selling Losers: implications for Stock Market Efficiency”, Journal of Finance, 48, 65-91 Jegadeesh, Narasimhan, and Sheridan Titman, 2001, “Profitability of Momentum Strategies: An Evaluation of Alternative Explanations”, Journal of Finance, 56, 699-720 Korajczyk, Robert A. And Roonie Sadka, 2004, “Are Momentum Profits Robust to Trading Costs?” Merton, R. C., 1974, “On the Pricing of Corporate Debt: The Risk Structure of Interest Rates”, Journal of Finance, 29, 449-70 Rouwenhorst, K. Geert, 1998, “International Momentum Strategies”, Journal of Finance, 53, 267-284 Rouwenhorst, K. Geert, 1999, “Local Return Factors and Turnover In Emerging Markets”, Journal of Finance, 54, 1439-1464 Shleifer, Andrei, and Vishny Robert W. 1997, “The limits of arbitrage”, Journal of Finance 52, 35–55 Zhang, X. Frank, 2006, “Information uncertainty and stock returns”, Journal of Finance 61, 105–136 35 Appendix Appendix A: Classification Function Coeficients of Discriminant Analysis Variable NPG PG Size 4.424 0.300 85.380 2.841 1.059 1.198 4.149 0.558 -30.859 3.626 0.301 89.630 6.830 2.165 1.703 4.978 1.216 -24.075 BTM Spd NegNI LG HG LP HP Constant In this appendix we show the coefficients of Fisher’s Linear Discriminant Function. These are the functions that lead to the classification results of Panel J Appendix B: Logit Regression without Dummy Variables Variable Coeficient C 1.726698 (26.45)*** Size (t-6) -0.300408 (-53.26)*** BTM (t-6) 0.020500 (4.94)*** Spd (t-6) 0.069270 (0.40) CR (t-6) -0.133283 (-2.55)** ROE (t-6) -0.000143 (-0.29) Observations 85068 LR Statistic 5527.154*** McFadden R² 0.067501 Log likelihood -38177.77 This panel shows the results of Logit Regression replacing Dummy variables for CR and ROE. As in panel L, in brackets we have z-statistics and the corresponding level of significance.