Interpreting qq-plot generated in R with a small sample of data I have a small data sample with 10 firms and want to specify the distribution of their abnormal returns.
First thing I did was to create the following histogram which indicates, when i am not mistaken that the data might somehow be normal distributed but is left skewed, right?

Another test I made to get a visuale representation of the distribution is a qq-plot:

Based on the qq-plot I would not assume the data to be normally distributed or does the fitting of the dots on the line is somehow good enough to state the opposit?
 A: Your data are definitely left skewed, but fairly close to a normal distribution. Both plots show the same thing in two different ways. 
The first plot shows you the distribution of your residuals as a histogram: the distribution of these points is not quite symmetric. You can try the normcheck function in the s20x package if you want a slightly cleaner visualisation of this that overlays the theoretical normal distribution.
The second plot with the dots and lines is slightly more informative but less intuitive. The line allows you to judge how close your observations (the dots) are to the theoretical normal distribution. The fact that some of the dots lie below the line on the lefthand side suggests that some of the points are further left than one would expect under a normal distribution given a large sample size. Ideally, all the points lie on the line.
Depending on your reasons for fitting the model a slight left skew like you have may not be a problem. If you are fitting the model, determine if there is a relationship between some covariates and a response. Then you can rely on the "central limit theorem", which basically says that given enough data points the underlying distribution will tend towards normal - i.e. don't worry about it in this case.
However, if you are building the model for prediction you are in a little bit of a predicament as the predictions are no longer going to be quite accurate. This is because the assumption that your errors are normally distributed has not been met. You can still salvage this though; you will just need to try a different model. Here are a few options:


*

*Log your response variable.

*Try a robust regression technique that is less susceptible to these issues.

*Quantile regression which makes no assumptions about the underlying error distribution.
