This research is an attempt to develop a conceptual framework for studying and modelling random variables in order to identify their analytical structures so as to provide adequate predictions of these variables. Such modelling can be useful in a number of instances. One of the uses of these models is to provide a better estimation of the causal econometric models involving stochastic regressors and the error term is likely to prevail. An illustration of these concepts was made on the basis of a time series of rainfall of the Kef region. The use of this climate variable is due to its twofold interests. Firstly, national planners use their forecast values to manage water resources. Secondly, it can be useful for better allocations of farmer's resources. Two methods were used to give some indications pertaining to these questions. The first one is mainly explicative in the sense that it provides a decomposition of the rainfall series structure into 4 components: trend, season, cycle and irregular factors. The second method used the ARIMA methodology and completes the first, in the sense that it develops a predictive tool for forecasting. The selected ARIMA model was used to develop an instrumental variable for the rainfall in the explanation of the variability in cereal yields in Tunisia. The results of the first method show the importance of the random part in the rainfall series. However, some non negligible components in series (trend, season, cycle) were identified and quantified. The second method enabled us to confirm the hypothesis that the utilization of the instrumental variables (ARIMA models) could improve the specification of the classical econometric models.