Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.

Winter Blues: A SAD Stock Market Cycle

1,406 views

Published on

Published in: Business, Economy & Finance
  • Be the first to comment

  • Be the first to like this

Winter Blues: A SAD Stock Market Cycle

  1. 1. Winter Blues: A SAD Stock Market Cycle By MARK J. K AMSTRA, LISA A. KRAMER, AND MAURICE D. LEVI* And God said, Let there be light; and tries to capture the in uence of daylight on there was light. And God saw that the human sentiment, risk tolerance, and hence light was good. stock returns. Our results strongly support a Genesis 1:3 SAD effect in the seasonal cycle of stock re- turns that is both signi cant and substantial, Depression has been linked with seasonal even after controlling for well-known market affective disorder (SAD), a condition that af- seasonals and other environmental factors. Pat- fects many people during the seasons of rela- terns at different latitudes and in both hemi- tively fewer hours of daylight. Experimental spheres provide compelling evidence of a link research in psychology has documented a clear between seasonal depression and seasonal vari- link between depression and lowered risk- ation in stock returns: higher-latitude markets taking behavior in a wide range of settings, show more pronounced SAD effects and results including those of a nancial nature. Through in the Southern Hemisphere are six months out the links between SAD and depression and be- of phase, as are the seasons. tween depression and risk aversion, seasonal The remainder of the paper is organized as variation in length of day can translate into follows. In Section I, we discuss SAD, depres- seasonal variation in equity returns. Based on sion, and equilibrium market returns. In Section supportive evidence from psychology which II, we introduce the international data sets. In suggests SAD is linked closely with hours of Section III, we explain the construction of the daylight, we consider stock market index data variables intended to capture the in uence of from countries at various latitudes and on both SAD on the stock market. We document in sides of the equator. We model differences in Section IV the signi cance of the SAD effect, the seasonal variation of daylight across coun- both statistical and economic, and provide an example of the excess returns that arise from trading strategies based on the SAD effect. In * Kamstra: Atlanta Federal Reserve Bank, 1000 Peachtree Section V, we explore the robustness of the Street NE, Atlanta, GA 30309 (e-mail: mark.kamstra@atl. frb.org); Kramer: Rotman School of Management, Univer- SAD effect to changes in variable de nitions sity of Toronto, 105 St. George Street, Toronto, Ontario, and estimation methods. Section VI considers M5S 3E6, Canada (e-mail: lkramer@chass.utoronto.ca); SAD in the context of segmented versus inte- Levi: Faculty of Commerce and Business Administration, grated capital markets. Section VII concludes.1 University of British Columbia, 2053 Main Mall, Vancouver, BC V6T 1Z2, Canada (e-mail: maurice.levi@commerce. ubc.ca). We have bene ted from the suggestions of two I. SAD and the Stock Market anonymous referees of this Review, Stanley Coren, Rick Green, Steven Jones, Andrew Karolyi, George Kramer, Tim As John Keats has written, “Four seasons ll Loughran, Raj Mehra, Jacob Sagi, Bob Shiller, Dick Thaler, the measure of the year. There are four seasons participants at the meetings of the American Finance Asso- ciation, the Canadian Econometrics Study Group, the Ca- in the mind of man.” One important aspect of nadian Economics Association, the Scottish Institute for the seasons as they affect the mind is the re- Research in Investment and Finance, and seminar partici- duced daylight hours during the fall and winter pants at the following universities: Guelph, Manchester/ months. According to Norman E. Rosenthal UMIST, McMaster, Montreal, Notre Dame, San Francisco, ´ Toronto, Wilfrid Laurier, and York. The authors gratefully (1998), the recurrent problems associated with acknowledge nancial support of the Social Sciences and Humanities Research Council of Canada and the research 1 assistance of Andy Bunkanwanicha and Yang Wu. Any An Appendix to this paper is available at www. remaining errors are solely the responsibility of the authors. markkamstra.com or from the authors on request. The Ap- The views expressed here are those of the authors and not pendix provides detailed estimation results from Kamstra et necessarily those of the Atlanta Federal Reserve Bank or the al. (2000) as well as further robustness and sensitivity Federal Reserve System. checks described but not provided below. 324
  2. 2. VOL. 93 NO. 1 KAMSTRA ET AL.: WINTER BLUES: A SAD STOCK MARKET CYCLE 325 diminished daylight take on a particularly se- taking tendency in the form of a scale of vere form among the approximately 10 million “sensation-seeking” propensity, the most widely Americans who are af icted with seasonal af- used of which was developed by Marvin Zuck- fective disorder, where “affective” means emo- erman (1984). The scale is then correlated with tional. A further 15 million suffer a milder form, various biological and psychological phenom- “winter blues.” The problem is also extensively ena, as shown, for example, by Zuckerman et al. documented outside the United States, with (1980). Sensation-seeking measures used to similar proportions of sufferers in countries judge the propensity to take risk have been around the world.2 Jeanne Molin et al. (1996) extensively documented as reliable measures of and Michael A. Young et al. (1997) provide risk-taking tendency in nancial decision- evidence that seasonal depression is related to making settings by Alan Wong and Bernardo hours of daylight, and hence the effects of SAD Carducci (1991), Paula Horvath and Zuckerman may be more pronounced in countries at more (1993), and Howard Tokunaga (1993), among extreme latitudes where winter and fall days are others. When the willingness to take risk is relatively shorter. related to measured levels of anxiety and de- SAD is clinically de ned as a form of major pression, there is a distinct tendency for greater depressive disorder.3 While usually described in anxiety or depression to be associated with re- terms of prolonged periods of sadness and pro- duced sensation seeking and reduced general found, chronic fatigue, evidence suggests that willingness to take risk, as shown by Zucker- SAD is connected to serotonin dysregulation in man (1984, 1994), Gregory A. Marvel and Bar- the brain. Furthermore, positron emission to- bara R. Hartman (1986), Solange Carton et al. mography (PET) scans reveal abnormalities in (1992), and Carton et al. (1995). In a further the prefrontal and parietal cortex areas due to study of depression and risk aversion, Amy E. diminished daylight, as described in the Na- Eisenberg et al. (1998) conducted experiments tional Institute of Mental Health study by in which subjects differing in degree of depres- Robert M. Cohen et al. (1992). That is, there sion were faced with a series of choices between appears to be a physiological source to the de- pairs of risky and safe options, including some pression related to shorter days. SAD symptoms of a nancial nature. By setting the choices such include dif culty concentrating, loss of interest that in some cases the risky option was the in sex, social withdrawal, loss of energy, leth- default (not requiring action) and in other cases argy, sleep disturbance, and carbohydrate or the safe option was the default, the researchers sugar craving often accompanied by weight were able to distinguish risk aversion from pas- gain.4 For those affected, the annual onset of sivity: depressive symptoms correlated with SAD symptoms can occur as early as Septem- risk aversion. ber, around the time of autumn equinox. See The psychological studies cited above Steven C. Dilsaver (1990), for example. strongly support the view that the depression Experimental research in psychology has associated with shorter days translates into a documented a direct link between depression greater degree of risk aversion, leading to test- and heightened risk aversion. This link is estab- able hypotheses in the context of stock market lished by rst providing a measure of risk- returns. Those market participants directly af- fected by SAD can in uence overall market returns according to the well-established princi- 2 For example, the frequency of SAD in northern Canada ple that market equilibrium occurs at prices is documented by Robert J. Williams and G. G. Schmidt where marginal buyers are willing to exchange (1993). The incidence of SAD in Italy is discussed by with marginal sellers: aggregate demands and Gianni L. Faedda et al. (1993). There is even evidence, from the Mayo Clinic (2002), that SAD occurs in countries as supplies for risky versus riskless assets can close to the equator as India. These and others studies thereby affect equilibrium risk premia.5 The suggest approximately 10 percent of people suffer from SAD. 3 See, for example, David H. Avery et al. (1993), Ybe 5 Meesters et al. (1993), Georg Leonhardt et al. (1994), and R. For further details on the impact of the marginal trader Michael Bagby et al. (1996). on market equilibria, see the classic papers by John R. Hicks 4 Mayo Clinic (2002). (1963) and Gerald O. Bierwag and M. A. Grove (1965), as
  3. 3. 326 THE AMERICAN ECONOMIC REVIEW MARCH 2003 implication is a causal relationship between sea- percent, and nds support for a relation between sonal patterns in length of day and market returns. sunshine and market returns. Hirshleifer and Studies on individuals at extreme latitudes, Shumway (2003) present further evidence for a including work by Lawrence A. Palinkas et al. sunshine effect in a study of 26 international (1996) and Palinkas and Matt Houseal (2000), stock markets. Instead of studying sun or cloud, suggest the depressive effects of SAD and Melanie Cao and Jason Wei (2001) investigate hence risk aversion may be asymmetric about the in uence of temperature on stock market winter solstice. Thus two dates symmetric about returns and nd evidence of a link in eight winter solstice have the same length of night but international markets. All of these studies con- possibly different expected returns. We antici- sider weather at the level of cities in which the pate seeing unusually low returns before winter markets are located. solstice and abnormally high returns following Molin et al. (1996) show that among several winter solstice. Lower returns should com- environmental factors (minutes of sunshine, mence with autumn, as SAD-in uenced indi- length of day, temperature, cloud cover, precip- viduals begin shunning risk and rebalancing itation, global radiation, and barometric pres- their portfolios in favor of relatively safe assets. sure), length of day has the strongest correlation We expect this to be followed by abnormally with seasonal depression. They employ step- high returns when days begin to lengthen and wise regression to reduce their multiple regres- SAD-affected individuals begin resuming their sion model to include only length of day and risky holdings. As long as there are SAD suf- temperature from the set of all the environmen- ferers shunning risk at some times of the year tal variables considered. Young et al. (1997) relative to other times, market returns will con- provide further evidence that SAD is related to tain a seasonal. According to the medical evi- the length of day by studying latitude. Our study dence on the incidence of SAD, this seasonal builds on the psychology literature linking sea- relates to the length of the day, not to changes in sonal affective disorder to length of day as well the length of the day. Therefore, against the null as the economics literature linking environmen- hypothesis that there is no effect of the seasons tal factors to stock market returns. Thus we test related to SAD and the winter blues, our alter- for a relationship between length of night and native hypothesis is that seasonal depression stock returns, controlling for other environmen- brought on by short days lead to relatively lower tal factors which may in uence returns, includ- returns in the fall and relatively higher returns in ing cloud cover, precipitation, and temperature. the winter. II. Market Returns Data A. Length of Night and Other Environmental Factors The daily stock index return data used in this study are outlined in Table 1: four indices A related literature in economics investigates from the United States as well as indices from the in uence of weather on market returns. As eight other countries, chosen to represent large- argued by Edward M. Saunders (1993) and capitalization, broad-based economies at differ- David Hirshleifer and Tyler Shumway (2003), ent latitudes in both hemispheres.6,7 We include the number of hours of sunshine affects peo- ples’ moods and hence also possibly market returns. The amount of sunshine is affected by 6 Our selection of indices was dictated by several crite- cloud cover as well as the number of hours of ria, including the availability of a suf ciently long time daylight, and indeed, Saunders (1993) uses a series, the absence of hyperin ation, large capitalization, and representation of a broad range of sectors. measure of cloudiness by classifying the degree 7 All of our foreign stock market returns were obtained of cloudiness in New York City into three cat- from Datastream. A feature of the Datastream time series is egories: 0 –30 percent; 40 –70 percent; 80 –100 that they often include nontrading days such as holidays. Datastream typically assigns a value on a nontrading day equal to the previous day’s price or equivalently a zero return. To remove holidays, we made use of Datastream’s well as the Appendix, “The Equilibrium Prices of Financial vacation les (when available), which track various coun- Assets,” by James C. Van Horne (1984, pp. 70 –78). tries’ holidays (starting no earlier than 1985), augmented by
  4. 4. VOL. 93 NO. 1 KAMSTRA ET AL.: WINTER BLUES: A SAD STOCK MARKET CYCLE 327 TABLE 1—DAILY STOCK INDEX RETURN DATA WITH CORRESPONDING CITIES AND LATITUDES Country Index City Latitude United States S&P 500 New York 41°N United States NYSE New York 41°N United States NASDAQ New York 41°N United States AMEX New York 41°N Sweden Veckans Affarer ¨ Stockholm 59°N Britain FTSE 100 London 51°N Germany DAX 30 Frankfurt 50°N Canada TSE 300 Toronto 43°N New Zealand Capital 40 Auckland 37°S Japan NIKKEI 225 Tokyo 36°N Australia All Ordinaries Sydney 34°S South Africa Datastream Global Index Johannesburg 26°S Note: Latitudes are rounded to the nearest degree and re ect the location of the city corresponding to each index’s stock exchange. the largest exchange among the far northerly S&P 500 index. The rst column of statistics is markets (Stockholm, Sweden) and the largest the daily percentage mean return. The mean exchange in the Southern Hemisphere (Sydney, daily return is of the same order of magnitude Australia). Our longest time series is the U.S. for all of the indices, ranging from about 0.01 S&P 500, which spans over 70 years. The long- percent to 0.06 percent. Standard deviation of est spanning index we could obtain for South the daily returns varies across countries, with Africa is the Datastream Global Index of 70 South Africa being the most volatile (uncondi- large-cap stocks in that country. All of the in- tionally) at 1.34 percent and the U.S. NYSE and dices we consider are value-weighted returns AMEX the least volatile, at 0.84 percent. The excluding dividend payments. For the United largest single-day drop, a decline exceeding 28 States we also investigated CRSP equal- percent, was experienced in Australia during the weighted indices and CRSP indices of returns October 1987 crash. The largest single day including dividends, and as shown in the Ap- gains ranged from 7.60 percent to 15.37 per- pendix (available from the authors), we found cent. All of the return series are strongly skewed qualitatively identical results in all cases. The to negative returns, as is typical with stock relevance of market segmentation/integration market returns. All the return series are strongly across markets at different latitudes is addressed kurtotic as well. Conventional tests of normality in Section VI. (not reported) strongly reject the hypothesis that Table 2 displays simple summary statistics any of these return series are normally distributed, for the raw data used in this study, the daily as is also typical with stock market returns. percentage returns for the four indices in the The returns data are summarized in an aggre- United States, and an index for each of Sweden, gate form in Figures 1 and 2. Figure 1 plots the Britain, Germany, Canada, New Zealand, Ja- monthly means of the daily percentage returns pan, Australia, and South Africa. Directly be- averaged across all four of the United States low the name of each country is the period over indices. The graph shows that in September, the which the returns were collected. The sample month in which the onset of autumn occurs, sizes range from under 3,000 daily observations returns are on average at their lowest point of for New Zealand to over 19,000 for the U.S. the year. They gradually recover through the fall, and then in winter they become positive and peak in January, the rst month following information on foreign holidays gleaned from various winter solstice. Speaking very roughly there is sources including the Worldwide Holiday & Festival Site an indication in the data of some sort of cycle (www.holidayfestival.com). As shown in the Appendix (available from the authors), we nd qualitatively identical through fall and winter. The raw data suggest results whether controlling for holidays as described above that September and January may be extreme or omitting all zero return days. points on a seasonal cycle.
  5. 5. 328 THE AMERICAN ECONOMIC REVIEW MARCH 2003 TABLE 2—DAILY PERCENTAGE RETURN SUMMARY STATISTICS FOR EACH INDEX Country and period Mean Standard deviation Minimum Maximum Skew Kurtosis United States: S&P 500 0.024 1.12 220.47 15.37 20.35 18.90 1928/01/04–2000/12/29 United States: NYSE 0.035 0.84 218.36 8.79 21.16 28.76 1962/07/05–2000/12/29 United States: NASDAQ 0.047 1.10 211.35 10.57 20.48 12.08 1972/12/18–2000/12/29 United States: AMEX 0.032 0.84 212.75 10.56 20.86 16.41 1962/07/05–2000/12/29 Sweden 0.063 1.25 28.986 9.78 20.25 6.02 1982/09/15–2001/12/18 Britain 0.037 1.01 213.03 7.60 20.93 12.29 1984/01/04–2001/12/06 Germany 0.025 1.10 213.71 8.87 20.50 8.62 1965/01/05–2001/12/12 Canada 0.023 0.85 210.29 9.88 20.75 13.97 1969/01/03–2001/12/18 New Zealand 0.013 0.97 213.31 9.48 20.86 18.78 1991/07/02–2001/12/18 Japan 0.037 1.12 216.14 12.43 20.34 10.82 1950/04/05–2001/12/06 Australia 0.033 1.00 228.76 9.79 24.89 131.49 1980/01/03–2001/12/18 South Africa 0.054 1.34 214.53 13.57 20.72 9.69 1973/01/03–2001/12/06 Note: All indices are value-weighted and do not include dividend distributions. Figure 2 plots the monthly means of the daily turns for the United States indices. Panels A percentage returns averaged across the eight through D of Figure 4 plot the monthly means foreign indices. Prior to averaging the returns, for Sweden, Britain, Germany, and Canada, and the data for Australia, New Zealand, and South Panels A through D of Figure 5 plot monthly Africa were shifted six months to adjust for the mean returns for New Zealand, Japan, Austra- difference in seasons across the hemispheres. lia, and South Africa. (The horizontal axis of the Note that returns are not shifted in this manner plot for each country starts with the month in for any other gure or table in the paper. The which autumn equinox actually takes place in horizontal axis (in Figure 2 only) is labeled with that country, March in the Southern Hemisphere reference to timing in the Northern Hemisphere. and September in the Northern Hemisphere.) That is, the rst observation is an average of Each of the U.S. indices shown in Figure 3 dem- returns over the month in which autumn begins onstrate the same approximate pattern: lowest in each country, September for the Northern annual returns in the early fall months followed Hemisphere countries and March for the South- by increasing returns peaking in the rst month ern Hemisphere countries. In Figure 2, we ob- of the year. Similarly, for each of the countries serve the lowest average return in the month in shown in Figure 4, returns are at their lowest in which the onset of autumn occurs for all coun- September and then they peak shortly after win- tries, marked September, and the highest aver- ter solstice. The countries in Figure 5 are the age return occurs in the month following winter four closest to the equator among those we solstice, marked January. Overall the graph ex- consider, and thus we would expect to see less hibits what may be interpreted as a seasonal of a seasonal in these countries’ returns. While pattern similar to that shown in Figure 1. the pattern for Japan’s returns is similar to those Figures 3, 4, and 5 plot monthly means of we have observed for other countries, the pat- daily percentage returns for each of the individ- terns for South Africa, Australia, and New Zea- ual indices considered in this paper. Panels A land seem more random in nature. Regarding all through D of Figure 3 plot monthly mean re- the individual plots shown in Figures 3, 4, and
  6. 6. VOL. 93 NO. 1 KAMSTRA ET AL.: WINTER BLUES: A SAD STOCK MARKET CYCLE 329 FIGURE 1. COMPOSITE PLOT FOR THE FIGURE 2. COMPOSITE PLOT FOR FOREIGN INDICES UNITED STATES INDICES Notes: The Annual Mean represents the daily percentage Notes: The Annual Mean represents the daily percentage re- returns averaged over the year across all eight foreign turns averaged over the year across all four U.S. indices. The indices. The Monthly Mean represents the daily percentage Monthly Mean represents the daily percentage returns aver- returns averaged over each month across all eight foreign aged over each month across all seven value-weighted indices. indices. In producing this gure only, returns from indices in the Southern Hemisphere are shifted by six months to align the seasons, and then the horizontal axis is marked with reference to the timing in the Northern Hemisphere. 5, although the average annual patterns vary somewhat, they typically show weak mean re- turns in early autumn followed by strong returns Figure 6 shows the cycles for the length of the shortly after the longest night of the year. night for several of the countries included in this Broadly speaking, the returns drop thereafter study. For simplicity, the cycles shown in Fig- and atten out through the spring and summer. ure 6 re ect the length of night for the latitude The Southern Hemisphere exchanges, uncondi- at which a country’s stock exchange is located, tionally at least, do not follow the same pattern rounded to the nearest degree. The length of the in the fall, though perhaps it is not a surprise to night during the course of the year peaks in the nd a lack of seasonal patterns in the exchanges Northern Hemisphere on the winter solstice, located closest to the equator where seasonal December 21st, and reaches a trough on the uctuations in daylight are small. summer solstice, June 21st. 8 In the Southern Hemisphere, the function is six months out of III. Measuring the Effect of SAD phase, with its peak on June 21st and trough on December 21st. The variations in the length of In describing the seasonal pattern of light night are larger the further away one travels through the course of the year, one could equiv- alently consider the number of hours of day, from sunrise to sunset, or the number of hours 8 For convenience we assume winter solstice takes place of night, which simply equals 24 minus the on December 21 and summer solstice on June 21. In prac- number of hours of day. We choose the latter. tice the timing can vary by a couple of days.
  7. 7. 330 THE AMERICAN ECONOMIC REVIEW MARCH 2003 FIGURE 3. INDIVIDUAL PLOTS OF DATA FOR EACH OF THE U NITED STATES INDICES Notes: The Annual Mean for each index represents the daily percentage returns averaged over the year for that index. The Monthly Mean for each index represents the daily percentage returns averaged over each month for that index.
  8. 8. VOL. 93 NO. 1 KAMSTRA ET AL.: WINTER BLUES: A SAD STOCK MARKET CYCLE 331 FIGURE 4. INDIVIDUAL PLOTS OF DATA FOR SWEDEN, BRITAIN, G ERMANY, AND CANADA Notes: The Annual Mean for each index represents the daily percentage returns averaged over the year for that index. The Monthly Mean for each index represents the daily percentage returns averaged over each month for that index.
  9. 9. 332 THE AMERICAN ECONOMIC REVIEW MARCH 2003 FIGURE 5. INDIVIDUAL PLOTS OF DATA FOR NEW ZEALAND, JAPAN, AUSTRALIA, AND SOUTH AFRICA Notes: The Annual Mean for each index represents the daily percentage returns averaged over the year for that index. The Monthly Mean for each index represents the daily percentage returns averaged over each month for that index.
  10. 10. VOL. 93 NO. 1 KAMSTRA ET AL.: WINTER BLUES: A SAD STOCK MARKET CYCLE 333 months have relatively shorter days. Second, the measure varies similarly across entire coun- tries and even hemispheres; thus wherever a marginal trader happens to be located within a country, her degree of seasonal depression and hence her in uence on markets would be ex- pected to manifest similarly. A. SAD Measure Based on Normalized Hours of Night De ne H t as the time from sunset to sunrise at a particular location. Then we can de ne our SAD measure, SAD t, at that location as fol- lows9: (1) SAD t 5 5 H t 2 12 0 for trading days in the fall and winter otherwise. Note that SADt varies only over the fall and win- ter, the seasons when according to medical evi- dence SAD affects individuals. By deducting 12 (roughly the average number of hours of night FIGURE 6. HOURS OF NIGHT FOR SEVERAL MARKETS over the entire year at any location), SADt re ects Notes: Actual hours of night are shown for the latitudes at the length of the night in the fall and winter which various countries’ stock exchanges are located relative to the mean annual length of 12 hours. In (rounded to the nearest degree). The latitudes, in degrees, are as follows: 26 South for South Africa, 34 South for Sweden, for example, the SADt variable equals 0 Australia, 41 North for the United States, 50 North for at the autumn equinox (September 21), takes on Germany, and 59 North for Sweden. higher values until it peaks at 16 on winter sol- stice, then takes on lower values until it equals 0 at the spring equinox (March 20), and remains at 0 from the equator, i.e., the larger is the latitude through the spring and summer. For countries north or south. Thus, among the countries we closer to the equator, the value varies relatively consider, Sweden experiences much greater closer to 0 during the fall and winter. variability relative to countries closer to the The number of hours of night, H t, can be equator, like South Africa and Australia. determined using standard approximations from As described above, medical evidence indi- spherical trigonometry as follows. To calculate cates that variation in the amount of daylight the number of hours of night at latitude d we during the fall and winter has a systematic effect rst need the sun’s declination angle, l t: on individuals’ moods. Thus we use the number of hours of night during only the fall and winter (2) to capture the effects of SAD on markets. In Section V, we discuss the similarity of results that arise using alternate speci cations that allow l t 5 0.4102 z sin X D 2p 365 ~ julian t 2 80.25! for SAD-related effects through all four seasons. The length-of-night measure we use to cap- ture the effects of SAD on the stock market has 9 The fall and winter period is de ned as September 21 some desirable features. First, it allows us to see to March 20 for the Northern Hemisphere and March 21 to September 20 for the Southern Hemisphere. We assume the whether there are more pronounced stock mar- fall and spring equinoxes take place on September 21 and ket effects due to SAD in countries at more March 21, though the actual timing can vary by a couple of extreme latitudes where the fall and winter days.
  11. 11. 334 THE AMERICAN ECONOMIC REVIEW MARCH 2003 where julian t is a variable that ranges from 1 to 365 (366 in a leap year), representing the num- (4) D F all 5 t 5 1 for trading days in the fall 0 otherwise. ber of the day in the year. Julian t equals 1 for January 1, 2 for January 2, and so on. We can Including this dummy variable allows (but does then calculate the number of hours of night as: not require) the impact of SAD in the fall to differ from that in the winter. If, contrary to our (3) expectations, the effects are symmetric across the two periods, this will simply be re ected by 2 pd 24 2 7.72 z arccos 2tan 360 X D tan~ l t ! a coef cient on D Fall which is insigni cantly different from zero. t in the Northern Hemisphere X D Ht 5 2 pd 7.72 z arccos 2tan tan~ l t ! IV. In uence of the SAD Effect 360 in the Southern Hemisphere A. Estimation where arccos is the arc cosine. We run a single regression for each country, allowing the impact of SAD to vary freely from B. Asymmetry Around Winter Solstice country to country. Returns are regressed on up to two lagged returns (where necessary to con- There are at least two reasons to expect the trol for residual autocorrelation), a Monday length of night to lead to an asymmetric re- dummy, a dummy variable for a tax-loss selling sponse in market returns before winter solstice effect, the SAD measure, a fall dummy, cloud relative to after. First, as mentioned in Section I, cover, precipitation, and temperature.11 results from Palinkas et al. (1996) and Palinkas and Houseal (2000) indicate the depressive ef- (5) fect of SAD may be asymmetric around winter solstice. Second, the trading activity of SAD- r t 5 m 1 r 1 r t 2 1 1 r 2 r t 2 2 1 m Mo nday D Monday t affected investors may itself cause asymmetric patterns in equity returns. Speci cally, if inves- 1 m Tax D Ta x 1 m SAD SAD t 1 m Fall D F all t t tors become more risk averse at the onset of fall and then return to “normal” at the end of winter, 1 m Cloud Cloud t higher compensating returns for holding risky assets between those points in time are gener- 1 m Precipitation Precipitation t ated by an initial price which is lower than would otherwise have been observed. That is, 1 m Temperature Temperature t 1 « t . with the onset of heightened risk aversion asso- ciated with SAD, prices rise less quickly than they would otherwise. After levels of risk aver- 11 The use of the AR(1) speci cation for stock returns is sion return to their previous, non-SAD in u- common, as is the use of Monday and tax-loss dummy enced levels at the end of winter, the recovery variables. See, for instance, Vedat Akgiray (1989), Adrian of prices from their initial (lower) levels in- Pagan and William Schwert (1990), Raul Susmel and Rob- ert F. Engle (1994), and R. Glen Donaldson and Kamstra creases returns. The implication is that returns (1997). Seasonality in stock returns is explored in a wide are lower in the fall and higher in the winter. range of papers including Nai-Fu Chen et al. (1986), Eric C. In order to allow for an asymmetric effect in Chang and J. Michael Pinegar (1989, 1990), and Sven the fall relative to the winter, we introduce a Bouman and Ben Jacobsen (2002). There are numerous dummy variable for days of the year which are papers that study seasonal stock market effects as related to tax-loss selling, including Philip Brown et al. (1983), Seha in the fall season10: M. Tinic and Richard R. West (1984), Kiyoshi Kato and James S. Schallheim (1985), Richard H. Thaler (1987), Jay R. Ritter (1988), Steven L. Jones et al. (1991), Ravinder K. 10 Fall is de ned as September 21 to December 20 in the Bhardwaj and LeRoy D. Brooks (1992), George Athanas- Northern Hemisphere and March 21 to June 20 in the sakos and Jacques A. Schnabel (1994), Charles Kramer Southern Hemisphere. (1994), and James A. Ligon (1997).
  12. 12. VOL. 93 NO. 1 KAMSTRA ET AL.: WINTER BLUES: A SAD STOCK MARKET CYCLE 335 Variables are de ned as follows: r t is the period TABLE 3—AVERAGE ANNUAL PERCENTAGE RETURN DUE TO SAD AND DUE TO FALL D UMMY t return for a given country’s index, r t 2 1 and r t 2 2 are lagged dependent variables, D Monday is t Annual a dummy variable which equals one when pe- Annual return riod t is the trading day following a weekend return due to fall Unconditional (usually a Monday) and equals zero otherwise, Country due to SAD dummy annual return D Tax is a dummy variable which equals one for t United States: 9.2*** 23.6** 6.3*** a given country when period t is in the last S&P 500 trading day or rst ve trading days of the tax United States: 6.1* 22.5* 9.2*** year12 and equals zero otherwise, and D Fall is a NYSE United States: 17.5*** 28.1*** 12.5*** dummy variable which equals one for a given NASDAQ country when period t is in the fall and equals United States: 8.4*** 25.1*** 8.4*** zero otherwise. The environmental factors, each AMEX measured in the city of the exchange, are per- Sweden 13.5** 26.9** 17.1*** Britain 10.3** 22.3 9.6*** centage cloud cover (Cloud t), millimeters of Germany 8.2* 24.3** 6.5** precipitation (Precipitation t), and temperature Canada 13.2*** 24.3** 6.1*** in degrees Celsius ( Temperature t). 13 New Zealand 10.5** 26.6** 3.3 At the end of this subsection we discuss the Japan 6.9* 23.7** 9.7*** results from estimating equation (5). (Most im- Australia 5.7 0.5 8.8*** South Africa 17.5* 22.1 14.6*** portantly, the parameter estimates on the SAD variable and fall dummy variable will be shown Notes: In calculating the average annual return due to SAD to be statistically signi cant for almost all the for a particular country, we determine for each trading day indices we consider.) First we present Table the value of the SAD variable (which varies between zero and six for Sweden, for example, during the fall and winter, 3 which provides an analysis of each index’s and which equals zero otherwise), multiplied by that coun- average annual percentage return due to the try’s SAD variable estimate (from Table 4), then we adjust SAD variable and due to the fall variable. For the value to obtain an annualized return. Similarly, in cal- each index in our study we present in the rst culating the annual average return due to the fall dummy for column of statistics the average annualized re- a particular country, we determine for each trading day that country’s fall dummy variable estimate (from Table 4) mul- turn due to our SAD measure. In computing the tiplied by the value of the fall dummy variable (one during return due to SAD for a particular index, we the fall, zero otherwise), then we adjust the value to obtain calculate for each trading day the value of the an annualized return. In the case of the columns for the SAD variable (which varies between zero and annualized returns due to the SAD and fall dummy vari- ables, signi cance is based on t-tests on the parameter six for Sweden, for example, during the fall and estimates from Table 4. In the case of the unconditional winter, and which equals zero otherwise), mul- return column, signi cance is based on t-tests for a mean tiply by that index’s SAD variable estimate daily return different from zero. (presented in Tables 4A– 4C), and adjust the * Signi cant at the 10-percent level, one-sided. ** Signi cant at the 5-percent level, one-sided. *** Signi cant at the 1-percent level, one-sided. 12 According to Ernst & Young (1998), the tax year commences on January 1 in the United States, Canada, Germany, Japan, and Sweden. The tax year starts on April 6 in Britain, on July 1 in Australia, on March 1 in South value to obtain an annualized return. In the next Africa, and on April 1 in New Zealand. For Britain, since column we present for each exchange the aver- the tax year ends on April 5, the tax-year dummy equals one age annualized return due to the fall dummy for the last trading day before April 5 and the rst ve variable. In calculating the return due to the fall trading days starting on April 5 or immediately thereafter. dummy, we calculate for each trading day that The tax-year dummy is de ned analogously for the other countries in the sample. country’s fall dummy variable estimate (pre- 13 All of the climate data (cloud cover, precipitation, and sented in Tables 4A– 4C) multiplied by the temperature) were obtained from the Climate Data Library value of the fall dummy variable (one during the operated jointly by the International Research Institute for fall, zero otherwise), then we adjust the value to Climate Prediction and the Lamont-Doherty Earth Obser- vatory of Columbia University: ingrid.ldeo.columbia.edu. obtain an annualized return. The nal column We are grateful to these organizations for making the data presents the average annual percentage return available. for each index. One, two, and three asterisks
  13. 13. 336 THE AMERICAN ECONOMIC REVIEW MARCH 2003 indicate signi cance at the 10-, 5-, and 1- TABLE 4A—REGRESSION RESULTS FOR EACH OF THE U.S. INDICES percent levels, respectively, based on one-sided hypothesis tests. For the annualized returns due r t 5 m 1 r 1r t2 1 1 r 2r t2 2 1 m MondayD M day on t to the SAD variable and the fall dummy, signi - 1 m TaxD Tax 1 m SADSAD t 1 m FallD Fall t t cance is based on t-tests on the corresponding 1 m CloudCloudt 1 m PrecipitationPrecipitationt 1 m TemperatureTemperature t 1 « t coef cient estimates from estimating equation (5) (to be discussed below) upon which the calcula- Panel A: Parameter Estimates (Heteroskedasticity-robust t-tests) tions of the annualized returns are based. For the unconditional returns, signi cance pertains to t- S&P 500 NYSE NASDAQ AMEX Parameter 41°N 41°N 41°N 41°N tests on daily returns differing from zero. There are some striking aspects of Table m 20.049 0.015 20.017 0.051 (20.45) (0.13) (20.11) (0.47) 3. First, the average annualized return due to r1 0.063*** 0.151*** 0.145*** 0.269*** SAD is positive in all countries, ranging from (3.54) (5.93) (4.93) (8.80) r2 20.042*** — — — 5.7 percent to 17.5 percent. For the most part, (22.40) — — — countries at latitudes closer to the equator tend m Monday 20.209*** 20.124*** 20.256*** 20.280*** to have lower, less signi cant returns due to (29.50) (25.20) (27.50) (212.0) m Tax 0.065 0.010 0.067 0.183*** SAD than countries further from the equator. (1.14) (0.14) (0.72) (2.71) (In many cases, the return due to SAD exceeds m SAD 0.038*** 0.026* 0.071*** 0.035*** the entire unconditional annual return.) Second, (2.43) (1.61) (2.96) (2.34) m Fall 20.058** 20.040* 20.134*** 20.084*** the average annualized return due to the fall (22.20) (21.40) (23.30) (23.20) dummy is negative in all countries except for m Cloud 0.115 0.046 0.087 0.021 Australia. Third, the positive return due to SAD (0.74) (0.27) (0.40) (0.13) m Precipitation 20.002 20.001 20.003 20.002 combined with the negative return due to the (20.59) (20.21) (20.61) (20.84) fall dummy suggests that on balance the season- m Tem ture pera 0.003** ,0.001 0.003* 0.001 (1.80) (0.26) (1.35) (0.54) ally asymmetric effects of SAD are shifting returns from the fall to the winter. Panel B: Diagnostics Results from estimating equation (5) for each R 2 0.011 0.027 0.033 0.091 of our 12 indices appear in Tables 4A– 4C. AR(10) Names of all the parameters from equation (5) p-value 0.136 0.850 0.023 0.003 are indicated in the rst column, and estimates Notes: This table reports coef cient estimates from running for each market appear in the cells under each the indicated regression for each of the four indices consid- market’s name. Below each parameter estimate ered. Returns (r t) are regressed on a constant (m), lagged is a heteroskedasticity-robust t-statistic.14 In returns where necessary (one lag is required for all indices cases where a particular parameter was not es- except the S&P 500, which requires two), a dummy for the trading day following the weekend (D Monday), a dummy for timated (r1 and/or r2 for some indices), a dash t the last trading day and rst ve trading days of the tax year (—) appears. (We only used as many lagged Tax (D t ), the SAD measure (SAD t 5 H t 2 12 during fall depended variables as was required to eliminate and winter, 0 otherwise, where H t 5 number of hours of residual autocorrelation up to the 1-percent night), a dummy for trading days in the autumn (D Fall), t level of signi cance. Tests for autocorrelation percentage cloud cover (Cloud t), millimeters of precipita- tion (Precipitation t), and temperature in degrees Celsius are discussed below.) One, two, and three as- (Temperature t). All indices are value-weighted and do not terisks indicate signi cance at the 10-, 5-, and include dividend distributions. Beneath each index’s name, 1-percent levels, respectively, based on one- we indicate the latitude of the city in which the exchange is sided hypothesis tests. located. In Panel A, we present parameter estimates with associ- We see in Tables 4A–C that the SAD coef- ated t-statistics in parentheses immediately below (calcu- cient estimate is uniformly positive across all lated using heteroskedasticity-robust standard errors). For countries and is signi cant in all countries ex- cases where particular parameters were not estimated, the cept one. The fall dummy coef cient estimate is relevant cells contain a dash (—). In Panel B, we present the R 2 for each regression as well as the p-value for a x2 test for autocorrelation up to ten lags. 14 The standard errors are calculated using a modi cation * Signi cant at the 10-percent level, one-sided. of Halbert White’s (1980) standard error estimator sug- ** Signi cant at the 5-percent level, one-sided. gested by James G. MacKinnon and White (1985) and *** Signi cant at the 1-percent level, one-sided. recommended for its sampling distribution properties.
  14. 14. VOL. 93 NO. 1 KAMSTRA ET AL.: WINTER BLUES: A SAD STOCK MARKET CYCLE 337 TABLE 4B—REGRESSION RESULTS FOR EACH OF SWEDEN, TABLE 4C—REGRESSION RESULTS FOR EACH OF NEW BRITAIN, GERMANY, AND CANADA ZEALAND, JAPAN, AUSTRALIA, AND SOUTH AFRICA r t 5 m 1 r 1r t2 1 1 r 2r t2 2 1 m MondayD M day t on r t 5 m 1 r 1r t2 1 1 r 2r t2 2 1 m MondayD M day t on 1 m TaxD Tax 1 m SADSAD t 1 m FallD Fall t t 1 m TaxD Tax 1 m SADSAD t 1 m FallD Fall t t 1 m CloudCloudt 1 m PrecipitationPrecipitationt 1 m CloudCloudt 1 m PrecipitationPrecipitationt 1 m TemperatureTemperature t 1 « t 1 m TemperatureTemperature t 1 « t Panel A: Parameter Estimates Panel A: Parameter Estimates (Heteroskedasticity-robust t-tests) (Heteroskedasticity-robust t-tests) Sweden Britain Germany Canada South Parameter 59°N 51°N 50°N 43°N New Zealand Japan Australia Africa Parameter 37°S 36°N 34°S 26°S m 0.267* 0.231 0.097 20.067 (1.40) (1.36) (0.69) (20.50) m 20.400 0.003 0.106 20.273* r1 0.110*** 0.061* 0.057*** 0.153*** (20.73) (0.02) (0.48) (21.40) (3.94) (1.34) (3.04) (4.64) r1 — — 0.089** 0.088*** r2 — — — — — — (1.94) (3.63) — — — — r2 — — — — m Monday 20.036 20.117*** 20.149*** 20.131*** — — — — (20.76) (23.00) (24.80) (25.50) m Monday 20.204*** 20.040* 20.038 20.123*** m Tax 0.135 0.146* 0.165** 0.034 (23.80) (21.40) (21.10) (23.00) (0.84) (1.55) (1.67) (0.43) m Tax 20.218* 0.009 0.125* 0.012 m SAD 0.028** 0.030** 0.025* 0.052*** (1.50) (0.12) (1.60) (0.11) (1.97) (2.03) (1.58) (3.24) m SAD 0.047** 0.037* 0.029 0.113* m Fall 20.113** 20.036 20.070** 20.069** (1.70) (1.55) (0.89) (1.62) (22.00) (20.80) (21.90) (22.10) m Fall 20.108** 20.060** 0.007 20.033 m Cloud 20.374* 20.271 20.130 0.168 (22.20) (21.90) (0.22) (20.89) (21.30) (21.00) (20.53) (0.63) m Cloud 0.426 0.095 20.328 0.137 m Precipitation 0.001 20.017* 0.001 20.003 (0.56) (0.58) (21.00) (0.61) (0.09) (21.50) (0.10) (20.88) m Precipitation 0.001 20.004 20.005** 20.001 m Tem ture pera 20.001 20.001 0.001 0.002 (0.43) (21.10) (22.20) (20.26) (20.17) (20.30) (0.36) (1.16) m Tem ture pera 0.011* 20.001 0.005 0.015* (1.48) (20.47) (0.80) (1.56) Panel B: Diagnostics 2 Panel B: Diagnostics R 0.017 0.009 0.008 0.030 2 AR(10) R 0.010 0.002 0.010 0.010 p-value 0.131 0.218 0.029 0.019 AR(10) p-value 0.972 0.011 0.203 0.071 Notes: This table reports coef cient estimates from running the indicated regression for each of the four indices consid- Notes: This table reports coef cient estimates from running ered. Returns (r t) are regressed on a constant (m), lagged the indicated regression for each of the four indices consid- returns where necessary a dummy for the trading day fol- ered. Returns (r t) are regressed on a constant (m), lagged lowing the weekend (D Monday), a dummy for the last trad- t returns where necessary (New Zealand and Japan do not ing day and rst ve trading days of the tax year (D Tax), the t require lagged returns), a dummy for the trading day fol- SAD measure (SAD t 5 H t 2 12 during fall and winter, 0 lowing the weekend (D Monday), a dummy for the last trad- t otherwise, where H t 5 number of hours of night), a dummy ing day and rst ve trading days of the tax year (D Tax), the t for trading days in the autumn (D Fall), percentage cloud t SAD measure (SAD t 5 H t 2 12 during fall and winter, 0 cover (Cloud t ), millimeters of precipitation (Precipita- otherwise, where H t 5 number of hours of night), a dummy tion t), and temperature in degrees Celsius (Tempera- for trading days in the autumn (D Fall), percentage cloud t ture t). All indices are value-weighted and do not include cover (Cloud t ), millimeters of precipitation (Precipita- dividend distributions. Beneath each country’s name, we tion t), and temperature in degrees Celsius (Tempera- indicate the latitude of the city in which the exchange is ture t). All indices are value-weighted and do not include located. dividend distributions. Beneath each country’s name, we In Panel A, we present parameter estimates with associ- indicate the latitude of the city in which the exchange is ated t-statistics in parentheses immediately below (calcu- located. lated using heteroskedasticity-robust standard errors). For In Panel A, we present parameter estimates with associ- cases where particular parameters were not estimated, the ated t-statistics in parentheses immediately below (calcu- relevant cells contain a dash (—). In Panel B, we present the lated using heteroskedasticity-robust standard errors). For R 2 for each regression as well as the p-value for a x2 test for cases where particular parameters were not estimated, the autocorrelation up to ten lags. relevant cells contain a dash (—). In Panel B, we present the * Signi cant at the 10-percent level, one-sided. R 2 for each regression as well as the p-value for a x2 test for ** Signi cant at the 5-percent level, one-sided. autocorrelation up to ten lags. *** Signi cant at the 1-percent level, one-sided. * Signi cant at the 10-percent level, one-sided. ** Signi cant at the 5-percent level, one-sided. *** Signi cant at the 1-percent level, one-sided.
  15. 15. 338 THE AMERICAN ECONOMIC REVIEW MARCH 2003 negative in all countries except one and is sig- SAD portfolio allocation strategy in which she ni cantly negative in all cases except three. reallocated 100 percent of her portfolio twice a Overall, this is consistent with a SAD-induced year at fall and spring equinox, placing her seasonal pattern in returns as depressed and money in the Swedish market during the North- risk-averse investors shun risky assets in the fall ern Hemisphere’s fall and winter, then moving and resume their risky holdings in the winter, it into the Australian market for the Southern leading to returns in the fall which are lower Hemisphere’s fall and winter, her average an- than average and returns following the longest nual return would have been 21.1 percent night of the year which are higher than average. (which is 7.9 percent more than under the neu- Recall from Table 3 that the magnitude and tral strategy).15 By comparison, had she in- signi cance of the SAD effect is broadly related stead allocated her portfolio across countries in to latitude: countries at higher latitudes (where order to act against the SAD effect, moving her the seasonal variation in daylight is more ex- money into the Swedish market for the Northern treme) tend to experience larger, more signi - Hemisphere’s spring and summer and then into cant returns due to the SAD effect relative to the Australian market for the Southern Hemi- those closer to the equator. sphere’s spring and summer, her average annual Regarding other aspects of the estimation, we return would have been 5.2 percent (which is nd the Monday dummy and tax-loss dummy are 8.0 percent less than she would have earned signi cant for many countries. The parameter es- under the neutral strategy). timates associated with cloud cover, precipitation, Had the investor been willing and able to and temperature are typically insigni cant. In assume short positions, she might have at- Panel B of each table, we present the R2 for each tempted to pro t from the SAD effect by short- regression as well as the p-value for a x2 test ing the Swedish market during the Northern for autocorrelation up to ten lags. In all cases, we Hemisphere’s spring and summer and going fail to reject the null hypothesis of no residual long in the Australian market during the same autocorrelation at the 1-percent level, with the ex- time (when it is fall and winter in the Southern ception of the AMEX. An expanded model con- Hemisphere), then going long in the Swedish trolling for autocorrelation in the AMEX index market during the Northern Hemisphere’s fall produces qualitatively identical results. and winter while shorting the Australian mar- ket. The average annual return per dollar in- B. Trading Strategies vested in this strategy would have been 15.9 percent. Of course, had she taken long and short The results in Tables 3 and 4A– 4C suggest positions in the two markets in a reverse pattern, potential gains to employing a trading strategy acting against the SAD effect, her average an- based on the SAD effect. To explore this pos- nual return would have been 215.9 percent. sibility, we provide an illustration of the returns Overall, we nd that by adopting a pro-SAD earned by an investor from various trading strat- strategy, the investor in our example would egies for a sample pair of countries, one from have realized relatively substantial excess re- each hemisphere. We select for our example turns, while trading against the SAD effect Sweden and Australia, each being one of the would have been costly.16 Interestingly, for all most extremely located countries in its hemi- sphere, and each having available roughly the same length of returns data (starting in the early 15 The returns for all the trading strategies other than the 1980’s). Consider rst the benchmark “neutral” neutral strategy are based on the six-month returns corre- portfolio allocation strategy in which an inves- sponding to the fall and winter or the spring and summer, as appropriate, for the index in question. For simplicity, we tor in the early 1980’s placed 50 percent of her neglect transaction costs. portfolio in the Swedish index and 50 percent in 16 In considering pro-SAD trading strategies across other the Australian index. Twenty years later, the countries, we found the following. Applying the pro-SAD average annual return to this neutral strategy portfolio allocation strategy, an investor would have en- joyed positive excess returns reallocating her funds between would have been 13.2 percent (using annualized Australia and all Northern Hemisphere indices, as well as returns based on the daily returns shown in between New Zealand and all Northern Hemisphere mar- Table 2). Next, had the investor adopted a pro- kets except the United States. Pro-SAD trading strategies
  16. 16. VOL. 93 NO. 1 KAMSTRA ET AL.: WINTER BLUES: A SAD STOCK MARKET CYCLE 339 the countries we consider, the volatility of re- must take. Therefore, we considered some al- turns does not vary appreciably across the ternate measures including using the number of spring/summer and fall/winter periods, suggest- hours of night normalized to lie between 0 and ing that there is little difference in risk across 1, the number of hours of night normalized to the trading strategies we consider. lie between 21 and 11, and allowing the SAD measure to take on nonzero values through all four seasons. Broad qualitative statements that V. Robustness Checks can be made about the economically signi cant magnitude, sign, and statistical signi cance of All of the detailed estimation results de- the SAD effect through the fall and winter are scribed in this section are provided in the Ap- roughly the same across all measures we pendix (available from the authors). explored. A. Maximum-Likelihood Model C. SAD and Asymmetry Around Winter Solstice In the previous section, we addressed the issue of autocorrelation by introducing lags of All of the results reported above were based the dependent variable, and we addressed the on allowing for asymmetry around winter sol- possibility of heteroskedasticity by using stice by using a dummy variable set to equal one White (1980) standard errors. We have also during trading days in the fall. We also explored estimated a maximum-likelihood model an alternative parameterization, allowing the which jointly models the mean and variance fall to differ from the winter by splitting the of returns. We controlled for heteroskedasticity SAD measure into two separate regressors. using the Sign-GARCH model of Lawrence R. (One regressor was set to equal the value of the Glosten et al. (1993) and produced robust t-tests SAD measure during the fall and zero other- using the robust (to heteroskedasticity and non- wise. The other regressor was set to equal the normality) standard errors of Tim Bollerslev and value of the SAD measure during the winter and Jeffrey M. Wooldridge (1992).17 With the excep- zero otherwise.) There was no need for a fall tion of some minor quantitative changes, overall dummy variable in this case. We found no qual- the maximum-likelihood results are very similar itative difference in results to those presented in to the results reported above. Though we typically this paper. With both types of parameteriza- nd coef cients on the SAD variable and the tions, we nd evidence of the same asymmetric tax-loss and fall dummies are slightly reduced in SAD effects. magnitude, we do still see large, economically We additionally explored the timing of meaningful effects due to SAD. breakpoints (around which point in time the fall/winter asymmetry would revolve), as well B. Other Measures for SAD as whether we should allow for any asymmetry at all. Allowing for a breakpoint at winter sol- As previously discussed, theory does not stice seemed best able to capture the asymmetry specify the exact form that the SAD measure of the SAD effect. Results using different breakpoints or using no breakpoint produce broadly supportive results. would have yielded a mix of positive and negative returns for reallocations between South Africa and Northern Hemi- D. Rede ning the Tax-Loss Variable sphere markets, perhaps due to the fact that many large South African companies are gold-producing companies which are cross-listed in London and New York. Detailed In the results presented above, the tax-loss results on strategies based on all combinations of the dummy variable equals one for a given country markets we consider are available on the Web at www. when period t is in the last trading day or rst markkamstra.com or from the authors on request. ve trading days of the tax year and equals zero 17 The Sign-GARCH model was selected because among commonly applied methods, it tends to work most reliably. otherwise, consistent with what other studies in See, for example, Robert F. Engle and Victor K. Ng (1993) the tax-loss literature have documented. An- and Donaldson and Kamstra (1997). other possible speci cation is to de ne the
  17. 17. 340 THE AMERICAN ECONOMIC REVIEW MARCH 2003 tax-loss dummy to equal one for all the trading out of phase by six months relative to the north, days in the rst month of the tax year. When as expected. Overall, results are robust to dif- this alternate speci cation is employed, the ferent measures to capture the effect of SAD, SAD effect is somewhat stronger and the tax- and do not appear to be an artifact of heteroske- loss effect is somewhat weaker in signi cance. dastic patterns in stock returns. Supporting our argument is the fact that day- VI. Market Segmentation and SAD Effects light has been shown in numerous clinical stud- ies to have a profound effect on people’s To the extent that there is cross listing of moods, and in turn people’s moods have been stocks from, for example, Australia, that also found to be related to risk aversion. SAD is a trade in New York as American Depositary recognized clinical diagnosis, with recom- Receipts (ADRs) or in other forms, arbitrage mended treatments including light therapy, would tend to dampen any potential SAD ef- medication, and behavior modi cation: suffer- fects in the much smaller Southern Hemisphere ers are urged to spend time outdoors or take markets. Nevertheless, we still nd evidence of vacations where daylight and sunlight are more a SAD effect in the Southern Hemisphere de- plentiful. Of course, we are not suggesting these spite any dampening that might be occurring. treatments be applied to in uence market re- Also, if the international capital markets were turns. Rather, we believe that we have identi ed fully integrated, there would be dampening of another behavioral factor that should not be the SAD effect across the hemispheres: inves- ignored in explaining returns. tors from the Northern Hemisphere would buy in the Southern Hemisphere at the very time that REFERENCES those in that hemisphere were selling, and vice versa. However, the evidence, as described by Akgiray, Vedat. “Conditional Heteroscedasticity Maurice D. Levi (1997), Karen K. Lewis in Time Series of Stock Returns: Evidence (1999), and others, suggests that markets are not and Forecasts.” Journal of Business, January fully integrated; there is a strong home-equity 1989, 62(1), pp. 55– 80. bias. Market segmentation is also supported by Athanassakos, George and Schnabel, Jacques A. correlations between national savings and in- “Professional Portfolio Managers and the vestment rates as shown by Martin Feldstein January Effect: Theory and Evidence.” Re- and Charles Horioka (1980): these correlations view of Financial Economics, Fall 1994, 4(1), are higher than one would expect in an inte- pp. 79 –91. grated international capital market. Further- Avery, David H.; Bolte, Mary A.; Dager, Stephen more, even if international capital markets were R.; Wilson, Laurence G.; Weyer, Matthew; integrated, the dominant size of the Northern Cox, Gary B. and Dunner, David L. “Dawn Hemisphere markets would mean that we would Simulation Treatment of Winter Depression: still expect to see a SAD effect (albeit that of the A Controlled Study.” American Journal of Northern Hemisphere). Psychiatry, January 1993, 150(1), pp. 113– 17. VII. Conclusion Bagby, R. Michael; Schuller, Debra R.; Levitt, Anthony J.; Joffe, Russel T. and Harkness, The preponderance of the evidence in this Kate L. “Seasonal and Non-Seasonal Depres- paper supports the existence of an important sion and the Five-Factor Model of Personal- effect of seasonal affective disorder on stock ity.” Journal of Affective Disorders, June market returns around the world. Speci cally, 1996, 38(2–3), pp. 89–95. even when controlling for the in uence of other Bhardwaj, Ravinder K. and Brooks, LeRoy D. environmental factors and well-known market “The January Anomaly: Effects of Low Share seasonals, we still nd a large and signi cant Prices, Transactions Costs and Bid-Ask SAD effect in every northern country we con- Bias.” Journal of Finance, June 1992, 47(2), sider. In general the effect is greater the higher pp. 553–75. the latitude. Furthermore, evidence suggests the Bierwag, Gerald O. and Grove, M. A. “On impact of SAD in the Southern Hemisphere is Capital Asset Prices: Comment.” Journal
  18. 18. VOL. 93 NO. 1 KAMSTRA ET AL.: WINTER BLUES: A SAD STOCK MARKET CYCLE 341 of Finance, March 1965, 20(1), pp. 89 – Dilsaver, Steven C. “Onset of Winter Depression 93. Earlier than Generally Thought?” Journal of Bollerslev, Tim and Wooldridge, Jeffrey M. Clinical Psychiatry, June 1990, 51(6), p. 258. “Quasi-maximum Likelihood Estimation and Donaldson, R. Glen and Kamstra, Mark J. “An Inference in Dynamic Models with Time- Arti cial Neural Network—GARCH Model Varying Covariances.” Econometric Re- for International Stock Return Volatility.” views, 1992, 11(2), pp. 143–72. Journal of Empirical Finance, January 1997, Bouman, Sven and Jacobsen, Ben. “The Hallow- 4(1), pp. 17– 46. een Indicator, ‘Sell in May and Go Away’: Eisenberg, Amy E.; Baron, Jonathan and Selig- Another Puzzle.” American Economic Re- man, Martin E. P. “Individual Differences in view, December 2002, 92(5), pp. 1618–35. Risk Aversion and Anxiety.” Working paper, Brown, Philip; Keim, Donald B.; Kleidon, Allan Department of Psychology, University of W. and Marsh, Terry A. “Stock Return Sea- Pennsylvania, 1998. sonalities and the Tax-Loss Selling Hypoth- Engle, Robert F. and Ng, Victor K. “Measuring esis: Analysis of the Arguments and and Testing the Impact of News on Volatil- Australian Evidence.” Journal of Financial ity.” Journal of Finance, December 1993, Economics, June 1983, 12(1), pp. 105–27. 48(5), pp. 1749–78. Cao, Melanie and Wei, Jason. “Stock Market Ernst & Young International, Ltd. 1999 world- Returns: A Temperature Anomaly.” Working wide executive tax guide. New York: Ernst & paper, University of Toronto, November Young International, Ltd., September 1998. 2001. Faedda, Gianni L.; Tondo, Leonardo; Teicher, Carton, Solange; Jouvent, Roland; Bungener, M. H.; Baldessarini, Ross J.; Gelbard, H. A. Catherine and Widlocher, D. “Sensation Seek- ¨ and Floris, G. F. “Seasonal Mood Disorders, ing and Depressive Mood.” Personality and Patterns in Mania and Depression.” Archives Individual Differences, July 1992, 13(7), pp. of General Psychiatry, January 1993, 50(1), 843–49. pp. 17–23. Carton, Solange; Morand, Pauline; Bungener, Feldstein, Martin and Horioka, Charles. “Domes- Catherine and Jouvent, Roland. “Sensation- tic Saving and International Capital Flows.” Seeking and Emotional Disturbances in De- Economic Journal, June 1980, 90(358), pp. pression: Relationships and Evolution.” 314 –29. Journal of Affective Disorders, June 1995, Glosten, Lawrence R.; Jagannathan, Ravi and 34(3), pp. 219–25. Runkle, David E. “The Relationship Between Chang, Eric C. and Pinegar, J. Michael. “Sea- Expected Value and the Volatility of the sonal Fluctuations in Industrial Production Nominal Excess Return on Stocks.” Journal and Stock Market Seasonals.” Journal of Fi- of Finance, December 1993, 48(5), pp. nancial and Quantitative Analysis, March 1779– 801. 1989, 24(1), pp. 59–74. Hicks, John R. “Liquidity.” Economic Journal, . “Stock Market Seasonals and Pre- December 1963, 72(288), pp. 789– 802. speci ed Multifactor Pricing Relations.” Hirshleifer, David and Shumway, Tyler. “Good Journal of Financial and Quantitative Anal- Day Sunshine: Stock Returns and the ysis, December 1990, 25(4), pp. 517–33. Weather.” Journal of Finance, 2003. Chen, Nai-Fu; Roll, Richard and Ross, Stephen A. Horvath, Paula and Zuckerman, Marvin. “Sensa- “Economic Forces and the Stock Market.” tion Seeking, Risk Appraisal, and Risky Behav- Journal of Business, July 1986, 59(3), pp. ior.” Personality and Individual Differences, 383–403. January 1993, 14(1), pp. 41–52. Cohen, Robert M.; Gross, M.; Nordahl, Thomas Jones, Steven L.; Lee, Winston and Apenbrink, E.; Semple, W. E.; Oren, D. A. and Rosenthal, Rudolf. “New Evidence on the January Effect Norman E. “Preliminary Data on the Meta- before Income Taxes.” Journal of Finance, bolic Brain Pattern of Patients with Winter December 1991, 46(5), pp. 1909–24. Seasonal Affective Disorder.” Archives of Kamstra, Mark J.; Kramer, Lisa A. and Levi, General Psychiatry, July 1992, 49(7), pp. Maurice D. “Winter Blues: Seasonal Affec- 545–52. tive Disorder (SAD) and Stock Market
  19. 19. 342 THE AMERICAN ECONOMIC REVIEW MARCH 2003 Returns.” Working paper, University of Brit- Pagan, Adrian and Schwert, William. “Alterna- ish Columbia, January 2000. tive Models for Conditional Stock Volatil- Kato, Kiyoshi and Schallheim, James S. “Seasonal ity.” Journal of Econometrics, July 1990, and Size Anomalies in the Japanese Stock 45(1–2), pp. 267–90. Market.” Journal of Financial and Quantita- Palinkas, Lawrence A. and Houseal, Matt. “Stages tive Analysis, June 1985, 20(2), pp. 243–60. of Change in Mood and Behavior During a Kramer, Charles. “Macroeconomic Seasonality Winter in Antarctica.” Environment and Be- and the January Effect.” Journal of Finance, havior, January 2000, 32(1), pp. 128– 41. December 1994, 49(5), pp. 1883–91. Palinkas, Lawrence A.; Houseal, Matt and Leonhardt, Georg; Wirz-Justice, Anna; Krauchi, Rosenthal, Norman E. “Subsyndromal Sea- Kurt; Graw, Peter; Wunder, Dorothea and sonal Affective Disorder in Antarctica.” Haug, Hans-Joachim. “Long-Term Follow-up Journal of Nervous and Mental Disease, Sep- of Depression in Seasonal Affective Disor- tember 1996, 184(9), pp. 530–34. der.” Comparative Psychiatry, November– Ritter, Jay R. “The Buying and Selling Behavior December 1994, 35(6), pp. 457–64. of Individual Investors at the Turn of the Levi, Maurice D. “Are Capital Markets Interna- Year.” Journal of Finance, July 1988, 43(3), tionally Integrated?” in Thomas J. Courchene pp. 701–17. and Edwin H. Neave, Reforming the Cana- Rosenthal, Norman E. Winter blues: Seasonal dian nancial sector. Kingston, Ontario: affective disorder: What it is, and how to Queens University Press, 1997, pp. 63– 84. overcome it, 2nd Ed. New York: Guilford Lewis, Karen K. “Trying to Explain Home Bias Press, 1998. in Equities and Consumption.” Journal of Saunders, Edward M. “Stock Prices and Wall Economic Literature, June 1999, 37(2), pp. Street Weather.” American Economic Re- 571–608. view, December 1993, 83(5), pp. 1337–45. Ligon, James A. “A Simultaneous Test of Com- Susmel, Raul and Engle, Robert F. “Hourly Vol- peting Theories Regarding the January Ef- atility Spillovers between International Eq- fect.” Journal of Financial Research, Spring uity Markets.” Journal of International 1997, 20(1), pp. 13–32. Money and Finance, February 1994, 13(1), MacKinnon, James G. and White, Halbert. pp. 3–25. “Some Heteroskedasticity-Consistent Co- Thaler, Richard H. “Anomalies: The January variance Matrix Estimators with Improved Effect.” Journal of Economic Perspectives, Finite Sample Properties.” Journal of Econo- Summer 1987, 1(1), pp. 197–201. metrics, September 1985, 29(3), pp. 305–25. Tinic, Seha M. and West, Richard R. “Risk and Marvel, Gregory A. and Hartmann, Barbara R. Return: January vs. the Rest of the Year.” “An ‘Economic’ Theory of Addiction, Hypo- Journal of Financial Economics, December mania, and Sensation Seeking.” International 1984, 13(4), pp. 561–74. Journal of the Addictions, 1986, 21(4 –5), pp. Tokunaga, Howard. “The Use and Abuse of 495–508. Consumer Credit: Application of Psycholog- Mayo Clinic. Seasonal affective disorder. June ical Theory and Research.” Journal of Eco- 2002; online at www.mayoclinic.com. nomic Psychology, June 1993, 14(2), pp. Meesters, Ybe; Jansen, Jaap H.; Beersma, Dom- 285–316. ien G.; Bouhuys, A. L. Netty and van den Van Horne, James C. Financial market rates and Hoofdakker, Rudi H. “Early Light Treatment ows (2nd edition). Englewood Cliffs, NJ: Can Prevent an Emerging Winter Depression Prentice Hall, 1984. from Developing into a Full-Blown Depres- White, Halbert. “A Heteroskedasticity-Consis- sion.” Journal of Affective Disorders, Sep- tent Covariance Matrix Estimator and a Direct tember 1993, 29(1), pp. 41– 47. Test for Heteroskedasticity.” Econometrica, Molin, Jeanne; Mellerup, Erling; Bolwig, Tom; May 1980, 48(4), pp. 817–38. Scheike, Thomas and Dam, Henrik. “The In- Williams, Robert J. and Schmidt, G. G. “Fre- uence of Climate on Development of Win- quency of Seasonal Affective Disorder ter Depression.” Journal of Affective Among Individuals Seeking Treatment at a Disorders, April 1996, 37(2–3), pp. 151–55. Northern Canadian Medical Health Center.”
  20. 20. VOL. 93 NO. 1 KAMSTRA ET AL.: WINTER BLUES: A SAD STOCK MARKET CYCLE 343 Psychiatry Research, January 1993, 46(1), Zuckerman, Marvin. “Sensation Seeking: A pp. 41– 45. Comparative Approach to a Human Trait.” Wong, Alan and Carducci, Bernardo. “Sensation Behavioral and Brain Sciences, 1984, 7(3), Seeking and Financial Risk Taking in Every- pp. 413–71. day Money Matters.” Journal of Business and . Behavioral expression and biosocial Psychology, Summer 1991, 5(4), pp. 525–30. bases of sensation seeking. Cambridge: Cam- Young, Michael A.; Meaden, Patricia M.; Fogg, bridge University Press, 1994. Louis F.; Cherin, Eva A. and Eastman, Char- Zuckerman, Marvin; Buchsbaum, Monte S. and mane I. “Which Environmental Variables Are Murphy, Dennis L. “Sensation Seeking Related to the Onset of Seasonal Affective and Its Biological Correlates.” Psycholog- Disorder?” Journal of Abnormal Psychology, ical Bulletin, January 1980, 88(1), pp. 187– November 1997, 106(4), pp. 554– 62. 214.

×