Hi, I am doing a VAR project, but encountering a few questions. First, I have four variables, two of them are first difference stationary, but the other two are not stationary, and not sationary by first difference, second difference, taking log,and log difference. It seems that there is seasonality in the latter two variables. How to deal with this? I use RATS to do this. The RATS guide says that there is no need to make the variables stationary. Right? I am a bit doubting it. Thanks in advance for your advice!Mackinn

What VAR stands for? Why dont you try to detrend the data?

No, you do not need to make the variables stationary unless you are trying to put this in some kind of ARIMA framework with the consequent confidence bands as the VAR. I don't anyone who does it that way. I have no idea what your data is but there are several approaches you can take. One obvious one if possible is just to do the distribution of how much the item has historically moved over 10 day periods in the past. Or, if its a more complex derivative, use the pricing formula with some kind of stochastic inputs representative of history or potential future based on current market. If there is some insidious seasonality (like with commodities) you can decompose with the Census X-11 technique. I have not used RATS for a very long time so I do not know if that function is in there directly, but it is easy to program. As I recall, basically remove the trend and business cycle and what is left is seasonality and randomness. Compare the seasons (months, quarters, or whatever) across years to see the average change from zero, standardize it, and those are the seasonality coefficients.

and when I mean take away the trend, you can just use a combination of multiple regression (on time, time squared, time cubed) and judgement (for business cycle elimination) and not trying to difference it for that.

Thanks for your help! First, my four variables are exchange rate, nominal interest rate (the two are first difference statioanary), domestic output series(strong seasonality and trend) yt, domestic price variable pt (typical upward trend). the data is Sweden's data. Second, VAR is vector autoregression. Need the variables be made stationary before analysis? Third, If I want to make yt and pt stationary, how to deal with them? Thanks!Mackinn

I thought you meant Value at Risk. Vector Autoregression? I helped Thomas Sargent of VAR fame in macroeconomics quite a while ago but do not remember the details with enough clarity to offer advice. Maybe someone else could help.

Thats the reason why i asked him SanFranCA2002 All 4 variables must be stationary!!!Detrend and differentiate untill they become stationary. Use Unit Root tests as well

QuoteOriginally posted by: MackinnHi, I am doing a VAR project, but encountering a few questions. First, I have four variables, two of them are first difference stationary, but the other two are not stationary, and not sationary by first difference, second difference, taking log,and log difference. It seems that there is seasonality in the latter two variables. How to deal with this? I use RATS to do this. The RATS guide says that there is no need to make the variables stationary. Right? I am a bit doubting it. Thanks in advance for your advice!MackinnVariables don't have to be stationary in order to perform VAR analysis. However, they do have to be integrated of the same order. You seem to have a range of variables of different orders, so I doubt you can perform VAR analysis on them.VAR is based on the idea of cointegration. If a long run equilibrium relationship (a cointegrating relationship) exits between two variables, then the errors must be stationary. If one variable is stationary, and the other variable has a trend then there cannot be any equilibrium relationship between them.

Yeah, the variables I use seem to have different orders, the price index pt, is stationary with first difference, and has order 11 using Newy-west method, and the GDP yt is stationary with first difference, but has order 2, the exchange rate et is stationary with first difference and has order 4, interest rate is nearly stationary with first difference and has order 2. With the same method to choose order for the four variables, the order is 3. So according what you said, I will not be able to use VAR to model ?

QuoteOriginally posted by: MrQwertyVariables don't have to be stationary in order to perform VAR analysis. However, they do have to be integrated of the same order. You seem to have a range of variables of different orders, so I doubt you can perform VAR analysis on them.VAR is based on the idea of cointegration. If a long run equilibrium relationship (a cointegrating relationship) exits between two variables, then the errors must be stationary. If one variable is stationary, and the other variable has a trend then there cannot be any equilibrium relationship between them.Either i am completely wrong and i must burn my degrees and commit suicide as a punishment for all those years i ve spent studing without learning something usuful, or you are confusing the Vector AutoRegressive models with Vector Error Correction!! In a VAR model there are atleast 2 variables where we try to test if there is any dynamic interrelationship among them. In other words we are trying to test that given a set of variables x and y if a number of lagged values of y as well as the concurrent and a number of lagged values of the variable x can explain the current level of variable y and vice versa. I had employed VAR in the past to test if there are any interrelationships among the daily returns of some major European equity indices.Basically it can employed even in an excel spreadsheet since you know some basic matrix algebra to understand how the model worksThe Eviews package provides additionaly Granger causality tests, Impulse responses tests as well as Cholensky decomposition tests, to get a clearer picture of the existing dynamic interrelationships.Here i can guess that Mackin tries to test if there are any dynamic interrelationships among the interest rates FX rates GDP growth rates and Inflation. Mackin in your position i would employ the differences in interest rates, the returns of FX rates, and the differences of GDP growth and inflation rates. All of them are expected to be stationary, and the last two variables can proxy the "surprise" (that is unexpected changes, as opposed to the expected and already "diggested" information ) in GDP growth and inflation.The necessary number of lags for each variable will be given by Akaike information criterion (AIC). Schwartz Information Criterion seems to be more parsimonious.I hope this helpsRegardsAnthis

anyone know how to run a VAR in matlab. I get as far as setting up the specifications and am then not sure how to proceed?

Mackinn said:" am doing a VAR project, but encountering a few questions. First, I have four variables, two of them are first difference stationary, but the other two are not stationary, and not sationary by first difference, second difference, taking log,and log difference."If your variables are not stationary after 1st difference then you are in big mess. All Integrated and co-integrated variables theory are developed on assumption that system is of order of integration atmost 1. Because all our practical time series data can be fitted within this framework. For more analysis, see Lutkepohl. BTW may I know what time series data you are dealing with?Thanks,

QuoteI had employed VAR in the past to test if there are any interrelationships among the daily returns of some major European equity indices.I dont think your analysis was correct. Returns values are (atleast if it is daily return) Stationary/Stable. Therefore it is trivial that they are co-integrated. Actually if I strictly go with Co-integration definition, I should not say they are co-Integrated. I mean to say, as they are stationary, any linear combination of them will be stationary (because we are applying a stationary operator to stationary data). Therefore I think your analysis is something very trivial.Thanks,