Does the January Effect Still Exists?

The issue of the January Effect has attracted a lot of interest by both practitioners and researchers. The idea that stock returns in January are statistically bigger than in other months was first presented several decades ago. This study analyzes the issue of the January effect in a systematic and global way of studying the performance of 106 indexes in 86 countries and jurisdictions. It was observed that while this effect can still be appreciated in some markets it would appear that it is decreasing globally over time. It was also found that there appears to be an Inverted January Effect in several markets with the returns in January being lower than the returns in some other months. This analysis was performed with nonparametric tests. The hypothesis that the returns of the indexes do not follow in general a normal distribution was also confirmed with several tests.


Introduction
There are a large number of market abnormalities identified both in the academic literature as well as by practitioners. One of these abnormalities is the January effect. The January effect refers to the observation that returns in January appear to be higher than returns in other months. One of the firsts, if not the first, academic article describing the January Effect was (Watchel, 1942). Since then several authors, such as (Haugen, 1988), (Thaler, 1987), (Jones, 1989), (Moller, 2008), (He, 2011), have analyzedthis issue.The existence of a January effect, as many other market abnormalities, has been used as an argument supporting the idea that markets are not completely efficient. The idea behind this approach is that if such market abnormality exist and can be exploited for trading purposes then, at least in principle, it would be possible to outperform the market in a consistent way, which would contradict the market efficiency hypothesis. The scope of this article it is not to study the link between abnormal returns and market efficiencies but to analyze, in a global basis, what markets present this phenomenon.
The January effect has been observed in several countries for some specific time periods such as for instance the U.S. for some decades after World War I (Charles, 1989). The absence of a January effect before World War I was detected in other countries such as Germany (Taufiq, 2016). Interestingly, the results of (Taufiq, 2016) for the U.K. and the U.S. suggest that there was a January Effect in the pre-war period which seems to contradict several articles such as (Schultz, 1985). The majority of the literature available seems to support the hypothesis that there was no such effect for the period before the World War I, particularly in the U.S. Other countries where the January Effect has been detected are Japan (Li, 2015), Jordan, Morocco from 1988to 2014(Gharaibeh, 2017, Turkey (Guler, 2013), India (Kaur, 2017) and several countries in Western Europe (Asteriou, 2006(Asteriou, ) from 1991(Asteriou, to 2003 In some other markets, like Pakistan, an abnormal return in January has been identified (Hashmi, 2014) but the authors mentioned that the effect is small and no profitable strategy can be built after accounting for transaction costs. It is also interesting that the January effect seems to be changing over time with several articles, such as (Gu, 2013), (Mehdian, 2002), (Patel, 2016) pointing to a declining January effect in the U.S. market. In these articles the authors observed a decline in the effect in the U.S. market starting in the late eighties. The performance in January has been even treated as a precursor of the performance for the rest of the year (Cooper, 2006). There are also abundant articles defending the idea that there is no January effect in some markets such as New Zealand (Li, 2010), India (Mehta, 2009), (Pandeu, 2016) or Indonesia (Simbolon, 2015).The January Effect has been studied not only in equities but also in fixed income investments. For instance, (Starks, 2006) detected the presence of a January Effect on closed end municipal bond funds. Interestingly the authors detected the presence of a January effect on the funds but not on the bonds constituting these funds. The authors attribute these results to tax harvesting.

Data
The data are composed of monthly closing values of 106 indexes covering 86 countries and jurisdictions. It includes indexes representing supranational entities such as Europe or the GCC as well as special administrative areas such as Hong Kong in China. According to data from Bloomberg the combined market capitalization of those 84 countries and jurisdictions accounted for approximately 92.3% of the global market capitalization as of July 2017. There is no double counting, with the estimate excluding the market capitalization of supranational indexes such as those covering Europe or the GCC.
The length of the time series varies from country to country and from index to index. The Dow Jones index for instance has a much longer time series than some of the emerging markets indexes. The analysis was performed using the entire data set for each index as well as using, from comparability purposes, only the last 15 years of values as of end of June 2017. For consistency in all the cases the same numbers of data points per month were used. All the data were obtained from Bloomberg. Monthly returns were obtained using monthly closing prices and formula [1]. The data was then grouped by month (from January to December).

Procedure
In a preliminary test the hypothesis that the index returns are normally distributed were checked with a Lillie test and an Anderson Darling test for each index for every month. As expected, for most cases the hypothesis that the index returns are normally distributed was rejected at a 5% confidence level. The results of the Anderson-Darling and the Lillie tests for every month for every index can be found in Appendix 2 and Appendix 3. The null hypothesis of the Anderson Darling test is that the data come follow a normal distribution. For the vast majority of the indexes the hypothesis that the monthly returns follow a normal distribution (for all the months of the year) cannot be accepted. According to the Anderson Darling test there were only 4 indexes, out of the 106 analyzed, in which the assumption that the returns follow a normal distribution for all the 12 months of the year cannot be rejected. Those four indexes are the PSI All Share (Portugal), Nigerian Stock Exchange Index (Nigeria), Tunisian Stock Exchange Index (Tunisia) and the S&P NZX All Index (New Zealand). Using the Lillie test similar results were obtained with no rejection of the null hypothesis of a normal distribution only in 10 out of 106 indexes analyzed at a 5% significance level. The null hypothesis in the Lillie test is that the underlying data follows a normal distribution. The 10 indexes for which the hypothesis that their returns follow a normal distribution are the S&P 1500 (U.S.), Colombia Colcap Index (Colombia), Ibex 35 (Spain), PSI All Share All Share Index (Portugal), Oslo All Share Index (Norway), Vienna Stock Exchange Index (Austria), Tunisia Stock Exchange Index (Tunisia), Nigeria Stock Exchange Index (Nigeria), Tadawull All Share Index (Saudi Arabia) and Bloomberg GCC 200 (GCC).Given that for the vast majority of the indexes the monthly returns do not appear to follow normal distribution hence nonparametric tests, such as the Wilcoxon Rank Sum and the Kruskal Wallis tests, were used to compare the returns. These tests do not assume that the data follows a normal distribution. The Wilcoxon test compares the medians of two data sample to determine if they are statistically equal at a certain confidence level. The purpose of the Kruskal Wallis test is determining if two, or more, samples of data come from the same distribution or not at a determined confidence level.
The returns in January were compared with the returns for all the other eleven months of the year using the Wilcoxon test. The results for the Wilcoxon tests for the entire available data set for the data provider can be found in Table 2.
The results obtained using the Kruskal Wallis test can be found in Table 4. Given that the entire data series available for each index are not of the same size it seemed reasonable, for comparability purposes, to do some further analysis using the same data time period for all the indexes. The time period used was 15 years (ending in June 2017). It should be noted that not all the indexes have a time series of monthly returns for 15 years. In fact, of the 106 indexes analyzed 16 did not have data available for the required period. The list of indexes that did not fulfill this requirement can be seen in Table 1. The same process as before was repeated with these shorter times with a Wilcoxon and a Kruskal Wallis test performed in all of indexes. The results of the Wilcoxon test for this reduced data series can be found on Table 3  while the results of the Kruskal Wallis tests can be found on Table 4.

Results
The average returns per month for all the indexes analyzed as well as their standard deviations can be found in Appendix 1. These results, for an easier visualization, are presented grouped according to geographical characteristics and divided into five regions: Americas, Western Europe, Eastern Europe, Middle East & Africa and Asia & Oceania.

Wilcoxon (Entire Time Series)
Using the entire time series available in Bloomberg for the previously mentioned 106 indexes and after making the number of data points for each month equal a total of 66 indexes, representing 62.3% of the total, were found to have no statically different returns in the month of January compared to all the other 11 months of the year, 27 indexes had one month with returns statistically different from January, 8 with 2 months, 2 with 3 months, 2 with 4 months and 1 with 6 months. In most of the cases then there was no statistically significant difference. These results can be found in Table 2. The index with 6 months of statistically different returns compared to the returns of January was the OMX Tallin, an index representing the Estonian stock market. The median return, point estimate, for the OMX Tallin index is negative in January and statistically significantly lower than in March, April, July, August, September and November. The point estimates for the median in all these months were positive. These results would suggest that the returns in January for the OMX Index have been lower than the results in several other months. The January Effect is typically understood as the opposite effect with January having bigger returns than other months. The two indexes with 4 months statistically different from the results in January were the FTSE All Share Index (U.K.) and the BorsaIstambul 100 Index (Turkey). In both cases there was no significant difference in returns when comparing January to the rest of months in the first quarter. In the case of the U.K. index the point estimate for the median value of the return for the month of January was negative but small. The index had positive returns in all the 4 months with statistically significantly different returns (June, August, September and December). In the case of the Borsa Istambul Index the point estimates of the median followed a similar pattern than in the FTSE All Share index, with the same distribution of months (including the negative value for January and the positive value for all the other months), but with larger differences in the point estimated for the returns. The two indexes with 3 months with statistically different returns when compared to January are the Tanzanian All Share Index (Tanzania) and the Laos Composite Index (Laos). There is more disparity of results with these two cases. On one hand there is the case of the index for Tanzania. The results for this index were similar to the previous cases with the point estimate for the median coming negative but small in January followed by a few months of positive results (February, May and October). On the other hand there is the case of Laos, with rather strong results in January and three months of statistically significantly returns (April, September and December). It should be noted that the data series for the index is Laos is relatively short with only 6 years of returns available in Bloomberg (always fulfilling the requirement of having the same number of data points per month).
The eight indexes that have two months with returns statistically different from January can be seen in table 5. For none of the eight indexes there was a statistically significant difference between January and February. For a majority of indexes, 5 out of 8, there was a statistically significant difference between January and March. The results for the tests for these eight indexes (over the entire data series) seem to support that the results were statistically similar for most of the month. For the other two months the results in January seem to be weaker than in those two months for all these eight indexes. the results in February seem to be lower than the returns in January. Another four indexes had lower returns in April with the rest of months with lower returns distributed across the first half of the year. There were no months with significantly lower returns than January in the second half of the year i.e., after June.

Wilcoxon (Last 15 Years)
The first thing to notice is that, as previously mentioned, when performing the analysis on the data series including the last 15 years of returns there is a shorter number of indexes. This is due to the fact that not all the indexes have 15 years of returns available. For a list of the indexes excluded from this analysis please see Table 1. Of the 90 indexes examined 67, representing 74.4% of the total, presented no statistically significant different returns when using the Wilcoxon test. This represents a higher proportion that when analyzing the entire dataset available i.e., all the years, but excluding the indexes with less than 15 years of track record (the indexes in those two approaches would be the same but the time length of the data series would be different). When using the entire data series for the 90 indexes there is a total of 54 cases, representing approximately 60% of the cases, in which there is no statistically significant difference between the performance in January and the performance in any of the other months of the year. This seems to support the idea that the January-Effect might be dissipating over time.
When analyzing the last 15 years of data 17 indexes had returns statistically different in the month of January compared to another month in the same year. Three indexes had two months statistically different, two indexes three months and one index 7 months. The index that had 7 months with returns different from the returns on January was the Nigeria Stock Exchange Index. It would appear that in the case of Nigeria, during the last 15 years, that there are indications of a sizeable January Effect. The point estimate of the returns of this index is positive and statistically different from those on February, March, April, May, June, July and November with the majority of the point estimate for the returns in those months actually being negative.
The indexes that had three months of statisticallydifferent results were the OMX Tallim (Estonia) and the Malta Stock Exchange Index (Malta). Both of these indexes had negative median returns in the month of January. Similarly to the case when the entire data set is analyzed the Estonian index is one of the indexes that has the largest amount of months with statistically different returns when compared to the month of January. The returns in Estonian index seem to be negative for the month of January. It should be noted that the results in both cases, with the entire data set or only the last 15 years, the median of the returns is negative but the magnitude of the point estimate appears to decrease. This perhaps points to an inverse, but decreasing, January-Effect in the Estonian case. In the case of Malta the point estimate of the returns in January is also negative, with the months of March, August and September having statistically different results and positive point estimates for the median returns.
The indexes that have two months of statistically different results, compared to January, are the OMX Riga (Latvia), the Madex Casablanca (Morocco) and the Colcap Index (Colombia). In all these cases the point estimate of the median return for the month of January was negative or very close to zero while the point estimates for the months with statistically different returns were all positive. These results seem to indicate that there is no January-Effect in the traditional sense.
The 17 indexes that had one month of statistically different returns can be seen in table 7. Of all these 17 indexes the returns were higher in January, compared to the other statistical different month, for five indexes. These indexes are the Nasdaq 100 Index (U.S.), the Tel Aviv Stock Exchange Index (Israel) and the Blom Index (Lebanon), Philippines Stock Exchange Index (Philippines) and Sri Lanka Colombo Index (Sri Lanka). In most of these cases the months of lower returns followed closely after January. In the cases of theNasdaq 100 Index and the Tal Aviv Stock Exchange Index it occurred in February and in the cases of the Blom Index and the Philippines Stock Exchange Index in April. The exception of this trend was the Sri Lanka Colombo Index with the significantly different month happening in July. The point estimates of the median returns of all these three month were negative. There are no indications of higher returns in January for the other 14 indexes. Most of the statistically different performances happened in the first half of the year with only four cases happening after June. The four cases with different performances after June were the Bolsa de Valores de Panama Index (November), the OMX Stockholm All Index (October), Ukraine PFTS Index (October) and the Sri Lanka Colombo Index (July).

Kruskal-Wallis
The results for the Kruskal-Wallis test, for both the entire time series as well as the last 15 years, can be found in Table 1. This test tries to determine if the returns for all the 12 months come from the same distribution. Therefore one p value is obtained for all the 12 months rather than having 11 p values for the comparison of the months from February to December (comparing them to the returns in January).
At a 5% confidence level an according to the results from the Kruskal-Wallis test 85.8% of the indexes , using the entire data set available, the returns for all the 12 months came from the same distribution. When the entire length of the time series is used but those indexes with less than 15 years of data are excluded then in 85.6% of the cases the data come from the same distribution. When only the last 15 years of data are used, and indexes with less than those 15 years of returns are included in the analysis, then 93.3% of the cases appear to come from the same distribution supporting the idea that the January-Effect is apparently dissipating over time. It should be noted that the test does not specifically compared the performance in January with the performance in the rest of months. The test analyzes the returns in all those months in its entirety. Using the entire time series the 15 indexes that do not appear to have the returns for all the months coming from the same distribution can be seen in Table 8. The 13 and 6 indexes for the entire time series length (excluding those indexes with less than 15 years of returns) as the case only including the last 15 years of returns respectively can be seen in Tables 9 and 10.