Via : Journal of Financial Economics (May 1996) Usually the first multivariate time series model that one comes across is a VAR model. It is a logical progression from modeling a univariate ARMA process. Most of the textbooks that introduce VAR start off with the Standard VAR and then go at length in to procedures such as estimating the parameters, hypothesis testing for the number of lags to consider, innovation accounting topics such as Impulse Response Decomposition, Forecast error variance decomposition.

This paper by Hasbrouck is about estimating trading costs from transaction prices. One of the classic models used for estimating trading costs is the Roll model. For a plain version of Roll model where the price increments are modeled in a univariate sense, an estimate for the costs is given by a formula that involves square root of negative auto correlation. In cases where there is a positive autocorrelation between the transaction prices, the formula loses its power.

Here is one of the most cited papers in sociology, that is just 1.5 pages long. Good things come in small packages Choosing models for cross-classifications(Raftery, A.E. (1986)).

This short article by George Casella and Edward George, explains the nuts and bolts of a Gibbs sampler and answers the following questions in simple words : What is a Gibbs sampler ? Why was a there a need for such a sampling algorithm ? When is it used ? Why does it work ? How is it related to Data Augmentation Algorithm ? When does the algorithm fail ?

[youtube https://www.youtube.com/watch?v=DXrs7QVkT-w?rel=0]