Forecasting CPU availability in volunteer computing systems using a single prediction algorithm is insufficient due to the diversity of the world-wide distributed resources. In this paper, we draw-up the main guidelines to develop an appropriate CPU availability prediction system for such computing infrastructures. To reduce solution time and to enhance precision, we use simple prediction techniques, precisely vector autoregressive models and a tendency-based technique. We propose a predictor construction process which automatically checks assumptions of vector autoregressive models in time series. Three different past analyses are performed. For a given volunteer resource, the proposed prediction system selects the appropriate predictor using the multi-state based prediction technique. Then, it uses the selected predictor to forecast CPU availability indicators. We evaluated our prediction system using real traces of more than 226000 hosts of Seti@home. We found that the proposed prediction system improves the prediction accuracy by around 24%.