Abstract: Given the fact that atmosphere is by no means a determinate system，this paper considers the numerical weather forecast in the view of information theory．The initial value and boundary value can be regarded as input information(information sources and a numerical model is just a tool which exchanges information．Numerical models can transform input information into results of the prediction of future weather information．So the forecast accuracy is enslaved to two aspects：one is how much output information is contained in input information，the other is how much information will be lost in the process of information transformation．The process of generating initial value means that the observational data of a moment does not include all required information for a model's initial value，and the absent information is more or less hidden in the historical observations．Therefore，it is necessary to utilize historical data to increase predicted information which is included in the input data．This paper concludes that a numerical model which can generate more output information than input information，does not exist，and inreasing input information is of essential sense．Furthermore，this paper demonstrates that model errors are also more or less hidden in historical data．In order to make good use of the historical observational data，this paper suggests that the forecasting problem should be regarded as an inverse problem rather than an initial value problem．Data assimilation is essentially an inverse problem，and its under determination should not be artificially exaggerated．The numerical weather prediction，an inverse problem，can not only make full use of historical data，but also use synoptic methods，statistical methods and dynamical methods in combination．This inverse problem can be resolved in practice by a specific method which synthesizes distinction between operational analysis and research results，universality of models，as well as statistics of pertinence．Therefore，it is a feasible approach to use historical data to improve model predictions without constructing a new model，which by any means is a very difficult work．