Abstract: Combining the two streams of thoughts, i. e. the state-space reconstructed from single variable time series and traditional multi-variable analysis, a multi-variable state-space forecasting method was developed and applied to dekad rainfall prediction for two regions in eastern China. The new method, by considering temperature and seasonality in rainfall-state-space, improved the prediction correlation by 5% (for 437 pairs of data). The improvement remarkably deals with large rainfall deviations, thus having particular significance to meteorological prediction. Great potential of further improvement through combining more reasonable variables in state-space could be expected