time-series help

#1
Hi, I don't know much about time series, but so hoping someone here does that can help. I'm trying to build a model in SAS that forecasts the home price increases by city. My data has year-month and home price (Y variable) for the last 12 years by city. I also have some other variables like home start-ups, days-on-market, unemployment rates, income, and many others... can I use all these as X-variables in a time series model.. if so how in SAS? Thanks..
 

vinux

Dark Knight
#2
Selecting a time series model is based summary of exploratory analysis. One need to know the causal mechanism, trend/seasonality etc.

You could start with PROC REG, then PROC AUTOREG. REG will treat the data as cross section. In AUTOREG you can add X-variables are Auto regressive part.
 
#3
Thanks Vinux, very helpful.. Are you saying to first start with PROC REG, then AUTO REGn ad then somehow intergrate it into a Time Series after? But in PROC REG, I wouldn't have the benefit of seasonality and trend and time etc?
 

noetsi

Fortran must die
#4
SAS has an ARIMA module if you purchased it (it is not in the base SAS STAT it is in SAS ETS). That might be the best way to go, but ARIMA is very difficult and you have to know your data very well to use it.
 

vinux

Dark Knight
#5
Thanks Vinux, very helpful.. Are you saying to first start with PROC REG, then AUTO REGn ad then somehow intergrate it into a Time Series after? But in PROC REG, I wouldn't have the benefit of seasonality and trend and time etc?
No. PROC REG or PROC AUTOREG are two options. My approach is to build simple model first( I sometimes say, start with stupid model), then build the actual model. If all the assumptions of REG is holding good, then probably you don't need to go for AUTOREG. But if you see/consider significant autocorrelation in the residual, use AUTOREG.

TimeSeries regression is not considered easy. You need to ensure the matching part( i.e. it should not be the case that Y is seasonal and Xs are seasonal.. read first two pages of this). Covariate also time series, so strictly speaking, you need to build a VAR model. You need to take care Cointegration. I am not sure you are going that extend.

You could also use ARIMA, but you may not be able to include the covariates.
 

noetsi

Fortran must die
#6
A much simpler method to predict than ARIMA (based on what has occured in the past not covariates) is simple exponential smoothing which SAS also supports (in ETS).