One of the key variables for a bank's management is the development of the "risk-free interest rate", which is the reference for all bond and loan rates as well as an indicator for the state of the economy and therefore the bank"s future perspectives. Turning towards long-term analysis, the risk-free rate is usually supposed to be the return of a superior-rated government bond (in most cases the return of the German 10-year Government Bond). Due to the importance of this risk-free rate, nearly all large economic and financial institutions deal with the analysis of its future development. In this paper we try to find out whether modelling non-linear relationships between variables can enhance forecast ability. We apply multi-layer perceptrons (MLP) as non-linear modelling tool beside an error correction model and a basic structural model with ARMA terms. Using seasonally unadjusted monthly data from 1960-2003, we forecast the interest rate for a two year hold-out sample. The obtained results give evidence of the underlying non-linearity of the problem. The MLP outperform the classical tools with regard to different error measures.