I would like to model Interest Rates (e.g. Libor). I know about Vasicek, CIR, etc. models but it seems to me that this models are only suitable for pricing, not for any predictions. Am I right? If I can use them for modelling (predictions) not pricing, can I get their parameters (calibrate them) using only interst rates historical data (e.g. Libor historical data)? I have always found calibration via different products - bonds, options...Thanks in advance for any help.

It depends what you want to predict.If you just want a probability distribution of 1-month LIBOR one year from today, you don't need a complicated model. You could look at historical data, comparing 1-month LIBOR at time t with 1-month LIBOR and 12-months-forward 1-month LIBOR; both from a year earlier. If that wasn't accurate enough, you could bring in other explanatory variables.The trouble is if you want to predict combinations of interest rates, say 1-month and 3-month LIBOR one year from today; or the move in 1-month LIBOR 1 years from today versus 2 years from today; or the correlation between 1-month LIBOR moves and 3-month LIBOR moves. Then you need a model of interest rates.Generally you have a choice between models that are accurate about interest rate movements and models that give realistic yield curves. Depending on your application you might prefer one or the other.

you should try black-derman and toy model, it requires calibration to the current interest levels

what do you want to predict? the yield curve? no need to employ a term structure model, just look at fwd yields (good enough)

Actually I would like to use the models for VaR, more specifically - as a Government Bond Issuer - for estimating possible future cupon cost of the state debt. So far I have used historical simulation which gives good results but Interest Rates models seems to be more sophisticated.Than You for all Your replies.

we developed a real-world model for var type stuff. you can download the paper from www.rebonato.com

Most interest rate models do not give good distributions of future interest rates. As you say, they are designed for pricing, not prediction. I think either historical data or macro-economic simulation would give better results. Modeling would be useful as an adjunct to those things, but I wouldn't rely on it for the main calculation.