10
$\begingroup$

I am doing some descriptive statistics of daily returns on stock indexes. I.e. if $P_1$ and $P_2$ are the levels of the index on day 1 and day 2, respectively, then $log_e (\frac{P_2}{P_1})$ is the return I'm using (completely standard in literature).

So the kurtosis is huge in some of these. I'm looking at about 15 years of daily data (so around $260 * 15$ time series observations)

                      means     sds     mins    maxs     skews     kurts
ARGENTINA          -0.00031 0.00965 -0.33647 0.13976 -15.17454 499.20532
AUSTRIA             0.00003 0.00640 -0.03845 0.04621   0.19614   2.36104
CZECH.REPUBLIC      0.00008 0.00800 -0.08289 0.05236  -0.16920   5.73205
FINLAND             0.00005 0.00639 -0.03845 0.04622   0.19038   2.37008
HUNGARY            -0.00019 0.00880 -0.06301 0.05208  -0.10580   4.20463
IRELAND             0.00003 0.00641 -0.03842 0.04621   0.18937   2.35043
ROMANIA            -0.00041 0.00789 -0.14877 0.09353  -1.73314  44.87401
SWEDEN              0.00004 0.00766 -0.03552 0.05537   0.22299   3.52373
UNITED.KINGDOM      0.00001 0.00587 -0.03918 0.04473  -0.03052   4.23236
                   -0.00007 0.00745 -0.09124 0.06405  -1.82381  63.20596
AUSTRALIA           0.00009 0.00861 -0.08831 0.06702  -0.74937  11.80784
CHINA              -0.00002 0.00072 -0.40623 0.02031   6.26896 175.49667
HONG.KONG           0.00000 0.00031 -0.00237 0.00627   2.73415  56.18331
INDIA              -0.00011 0.00336 -0.03613 0.03063  -0.22301  10.12893
INDONESIA          -0.00031 0.01672 -0.24295 0.19268  -2.09577  54.57710
JAPAN               0.00008 0.00709 -0.03563 0.06591   0.57126   5.16182
MALAYSIA           -0.00003 0.00861 -0.35694 0.13379 -16.48773 809.07665

My question is: Is there any problem?

I want to do extensive time series analysis over this data - OLS and Quantile regression analysis, and also Granger Causality.

Both my response (dependent) and predictor (regressor) will have this property of gigantic kurtosis. So i'll have these return processes on either side of the regression equation. If the non-normality spills over into the disturbances that will only make my standard errors high variance right?

(Perhaps I need a skewness robust bootstrap?)

$\endgroup$
  • 3
    $\begingroup$ 1) You may want to move this to the quant.stackexchange.com site. 2) What do you mean by problem? There is a whole literature on the impact of outliers on moments. It can often be more of an art than a science. $\endgroup$ – John Aug 18 '12 at 5:20
  • 2
    $\begingroup$ "Is there any problem?" is too vague. What do you want to do with these data? Your huge kurtoses are associated with huge left skew. Since log(p2/p1) = log p2 - log p1, a huge left skew indicates that there were a few times when this was very low, that is, p1 much higher than p2, compared to the usual case. Could be a company going bankrupt or something like that. $\endgroup$ – Peter Flom Aug 18 '12 at 5:35
  • $\begingroup$ Sorry about this - I've amended my OP. $\endgroup$ – user13253 Aug 18 '12 at 5:44
  • 1
    $\begingroup$ log-returns are typically skewed and heavy tailed. For this reasons it is preferred to consider flexible distributions that can capture this behaviour. See for example 1 and 2. $\endgroup$ – user10525 Aug 18 '12 at 10:45
  • $\begingroup$ You should have a look at measures of kutosis based on L-moments $\endgroup$ – kjetil b halvorsen Sep 15 '15 at 14:01
2
$\begingroup$

Have a look at heavy-tail Lambert W x F or skewed Lambert W x F distributions a try (disclaimer: I am the author). In R they are implemented in the LambertW package.

Related posts:

One advantage over Cauchy or student-t distribution with fixed degrees of freedom is that the tail parameters can be estimated from the data -- so you can let the data decide what moments exist. Moreover the Lambert W x F framework allows you to transform your data and remove skewness / heavy-tails. Itt is important to note though that OLS does not require Normality of $y$ or $X$. However, for your EDA it might be worthwhile.

Here is an example of Lambert W x Gaussian estimates applied to equity fund returns.

library(fEcofin)
ret <- ts(equityFunds[, -1] * 100)
plot(ret)

Time series plot equity funds

The summary metrics of the returns are similar (not as extreme) as in OP's post.

data_metrics <- function(x) {
  c(mean = mean(x), sd = sd(x), min = min(x), max = max(x), 
    skewness = skewness(x), kurtosis = kurtosis(x))
}
ret.metrics <- t(apply(ret, 2, data_metrics))
ret.metrics

##          mean    sd    min   max skewness kurtosis
## EASTEU 0.1300 1.538 -18.42 12.38   -1.855    28.95
## LATAM  0.1206 1.468  -6.06  5.66   -0.434     4.21
## CHINA  0.0864 0.911  -4.71  4.27   -0.322     5.42
## INDIA  0.1515 1.502 -12.72 14.05   -0.505    15.22
## ENERGY 0.0997 1.187  -5.00  5.02   -0.271     4.48
## MINING 0.1315 1.394  -7.72  5.69   -0.692     5.64
## GOLD   0.1098 1.855 -10.14  6.99   -0.350     5.11
## WATER  0.0628 0.748  -5.07  3.72   -0.405     6.08

Most series show clearly non-Normal characteristics (strong skewness and/or large kurtosis). Let's Gaussianize each series using a heavy tailed Lambert W x Gaussian distribution (= Tukey's h) using a methods of moments estimator (IGMM).

library(LambertW)
ret.gauss <- Gaussianize(ret, type = "h", method = "IGMM")
colnames(ret.gauss) <- gsub(".X", "", colnames(ret.gauss))

plot(ts(ret.gauss))

time series plot of Gaussianized returns

The time series plots show much fewer tails and also more stable variation over time (not constant though). Computing the metrics again on the Gaussianized time series yields:

ret.gauss.metrics <- t(apply(ret.gauss, 2, data_metrics))
ret.gauss.metrics

##          mean    sd   min  max skewness kurtosis
## EASTEU 0.1663 0.962 -3.50 3.46   -0.193        3
## LATAM  0.1371 1.279 -3.91 3.93   -0.253        3
## CHINA  0.0933 0.734 -2.32 2.36   -0.102        3
## INDIA  0.1819 1.002 -3.35 3.78   -0.193        3
## ENERGY 0.1088 1.006 -3.03 3.18   -0.144        3
## MINING 0.1610 1.109 -3.55 3.34   -0.298        3
## GOLD   0.1241 1.537 -5.15 4.48   -0.123        3
## WATER  0.0704 0.607 -2.17 2.02   -0.157        3

The IGMM algorithm achieved exactly what it was set forth to do: transform the data to have kurtosis equal to $3$. Interestingly, all time series now have negative skewness, which is in line with most financial time series literature. Important to point out here that Gaussianize() operates only marginally, not jointly (analogously to scale()).

Simple bivariate regression

To consider the effect of Gaussianization on OLS consider predicting "EASTEU" return from "INDIA" returns and vice versa. Even though we are looking at same day returns between $r_{EASTEU, t}$ on $r_{INDIA,t}$ (no lagged variables), it still provides value for a stock market prediction given the 6h+ time difference between India and Europe.

layout(matrix(1:2, ncol = 2, byrow = TRUE))
plot(ret[, "INDIA"], ret[, "EASTEU"])
grid()
plot(ret.gauss[, "INDIA"], ret.gauss[, "EASTEU"])
grid()

scatterplot INDIA and EASTEU

The left scatterplot of the original series shows that the strong outliers did not occur at the same days, but at different times in India and Europe; other than that it is not clear if the data cloud in the center supports no correlation or negative/positive dependency. Since outliers strongly affect variance and correlation estimates, it is worthwhile to look at the dependency with heavy tails removed (right scatterplot). Here the patterns are much more clear and the positive relation between India and Eastern Europe market becomes apparent.

# try these models on your own
mod <- lm(EASTEU ~ INDIA * CHINA, data = ret)
mod.robust <- rlm(EASTEU ~ INDIA, data = ret)
mod.gauss <- lm(EASTEU ~ INDIA, data = ret.gauss)

summary(mod)
summary(mod.robust)
summary(mod.gauss)

Granger causality

A Granger causality test based on a $VAR(5)$ model (I use $p = 5$ to capture the week effect of daily trades) for "EASTEU" and "INDIA" rejects "no Granger causality" for either direction.

library(vars)  
mod.vars <- vars::VAR(ret[, c("EASTEU", "INDIA")], p = 5)
causality(mod.vars, "INDIA")$Granger


## 
##  Granger causality H0: INDIA do not Granger-cause EASTEU
## 
## data:  VAR object mod.vars
## F-Test = 3, df1 = 5, df2 = 3000, p-value = 0.02

causality(mod.vars, "EASTEU")$Granger
## 
##  Granger causality H0: EASTEU do not Granger-cause INDIA
## 
## data:  VAR object mod.vars
## F-Test = 4, df1 = 5, df2 = 3000, p-value = 0.003

However, for the Gaussianized data the answer is different! Here the test can not reject H0 that "INDIA does not Granger-cause EASTEU", but still rejects that "EASTEU does not Granger-cause INDIA". So the Gaussianized data supports the hypothesis that European markets drive markets in India the following day.

mod.vars.gauss <- vars::VAR(ret.gauss[, c("EASTEU", "INDIA")], p = 5)
causality(mod.vars.gauss, "INDIA")$Granger

## 
##  Granger causality H0: INDIA do not Granger-cause EASTEU
## 
## data:  VAR object mod.vars.gauss
## F-Test = 0.8, df1 = 5, df2 = 3000, p-value = 0.5

causality(mod.vars.gauss, "EASTEU")$Granger

## 
##  Granger causality H0: EASTEU do not Granger-cause INDIA
## 
## data:  VAR object mod.vars.gauss
## F-Test = 2, df1 = 5, df2 = 3000, p-value = 0.06

Note that it is not clear to me which one is the right answer (if any), but it's an interesting observation to make. Needless to say that this entire Causality testing is contingent on the $VAR(5)$ being the correct model -- which it is most likely not; but I think it serves well for illustratiton.

| cite | improve this answer | |
$\endgroup$
1
$\begingroup$

What is needed is a probability distribution model that better fits the data. Sometimes, there are no defined moments. One such distribution is the Cauchy distribution. Although the Cauchy distribution has a median as an expected value, there is no stable mean value, and no stable higher moments. What this means is that when one collects data, actual measurements crop up that look like outliers, but are actual measurements. For example, if one has two normal distributions F and G, with mean zero, and one divides F/G, the result will have no first moment and is a Cauchy distribution. So we happily collect data, and it looks OK like 5,3,9,6,2,4 and we calculate a mean that looks stable, then, all of a sudden we get a -32739876 value and our mean value becomes meaningless, but note, the median is 4, stable. Such it is with long-tailed distributions. Find a more correct long tailed distribution for your data, and use the statistical measurements that that distribution implies, and your problem will go away.

Edit: You might try Student's t-distribution with 2 degrees of freedom. That distribution has longer tails than the normal distribution, the skewness and kurtosis are unstable (Sic, do not exist), but the mean and variance are defined, i.e., are stable.

Next edit: One possibility might be to use Theil regression. Anyway, it's a thought, because Theil will work well no matter what the tails look like. Theil can be done MLR (multiple linear regression using median slopes). I have never done Theil for histogram data fitting. But, I have done Theil with a jackknife variant to establish confidence intervals. The advantage of doing that is that Theil doesn't care what the distribution shapes are, and, the answers are generally less biased than with OLS because typically OLS is used when there is problematic independent axis variance. Not that Theil is totally unbaised, it is median slope. The answers have a different meaning as well, it finds a better agreement between the dependent and independent variables where OLS finds the least error predictor of the dependent variable, which latter is not always the question that we want an answer to.

| cite | improve this answer | |
$\endgroup$
  • 2
    $\begingroup$ Nice info, thanks. Do you know some (quite compact) ressources to read further? I have a completely different problem with long tail, but I think my data is just a mixture distribution of different scenarios. $\endgroup$ – flaschenpost Jan 9 '16 at 22:18
  • $\begingroup$ I use Mathematica, and fitting distributions as well as defining distributions piece-wise is not difficult in that language. For example, look at this. In general, random variables add by convolution, but in practice convolution of density functions is challenging. Some people just piece-wise define density functions for admixed variables, for example adding a light exponential tail to a censored heavier gamma distribution after a maximum value to model earthquake frequency. @flaschenpost $\endgroup$ – Carl Jan 10 '16 at 0:25

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy