Forecasting the target variable vs building a causal model and forecasting causal variables I want to know the approaches people use to forecast lets say unemployment rate .... By itself it might not fit a time series model (ARMA) very well as the trend is dependent on many external factors. 
My hypothesis is that if a time series for the target doesn't fit well, but some of the underlying causal variables can be reasonably forecasted into the future it is better to use the forecasted causal variables and predict the target variable. Is that a common approach?
I don't know about ARIMA but does it take into account external factors?..
 A: No, this would not be a common approach to unemployment forecasting. The common approaches would be something like VAR, i.e. time series modeling. The field which deals with this stuff is called macroeconomic forecasting. 
The reason why particularly unemployment would be unsuitable for what you called causal forecasting is the endogeneity. Unemployment itself is such an important variable, that its state impacts the state of other variables. So, if you make model where you forecast macroeconomic variables, then use them to forecast unemployment, your model will be nonsensical, in my opinion. Unemployment, inflation and GDP must be in your core model in some form. It may not necessarily be unemployment rate, it could employment rate or some other form labor capacity utilization.
Having said that there are academic models which do not have unemployment in them, see e.g. the DSGE model: F. Smets and R. Wouters (2002), An Estimated Stochastic Dynamic General Equilibrium Model of the Euro Area, European Central Bank, Working Paper Series, No. 171. Also in Journal of the European Economic Association, Vol. 1, No. 5, 2003, pp. 1123-1175.
This MATLAB example will give you an idea of how typical forecasting models may look like. Note, they call their model DSGE, ignore it. It's not a DSGE model at all, this is a simple VAR model.
A: It seems like ARMAX/ARIMAX model can interest you. These models combine AR(p) with regression model.
If this is close enough to what you need, then yes this is pretty common approach.
