The development of a hydrological forecasting model is conditioned by the available data collected in the entire hydrological catchment for which discharge has to be predicted. The forecast quality not only depends on the available historical data but also on the continuously measured data that is necessary to update the model in real time according to the observed conditions. Consequently, the data type, their quality, reliability and the ease of data acquisition highly influence the final model choice. In Colombia, numerous catchments in the interior of the country are equipped with pluviometric and some of them with limnimetric stations but with often incomplete measurement series. For some catchments, temperature and evaporation data are also available but measured only on a sporadic basis. The acquisition of these data is often difficult, due to the accessibility, the size of the country and related costs. These factors have to be considered for the model development to ensure the use of representative and stable data. The present study analyses only one catchment, the one of the Rio Magdalena. It has been chosen for strategic and security reasons, in agreement with the persons in charge of the national hydrological service. For this catchment, the available data include water level measurements at different locations along the river (see the map in Section 3) and precipitation data measured at these limnimetric stations. The available stage-discharge relations are not reliable enough to build a discharge forecast model. A detailed measurement of the precipitation falling onto the catchment is difficult to carry out by conventional methods. The developed forecast models are of a so-called black-box type. They correspond to an autoregressive moving average model with exogenous variables. Each of them predicts river stages at a given point of the studied river. The explicative variables are either river stages or precipitation amounts, measured at the studied location or at other upstream stations. The forecast time step is one day and the forecasts are completed from one to five days ahead. The calibration of the model parameters is part of the model development process. For each limnimetric station, the resulting parameter values encode a model adapted to the available data and the considered forecast lag (time period between the forecast emission and the predicted event). In a first step, the developed models only include limnimetric data. In a second step, these models are further validated and affined considering pluviometric data. Models have been developed for all main stations of Rio Magdalena (see map at p. 24) but only the results for Barranca station are presented in details in this report. The obtained results can be considered as being good and reliable, as illustrated by the graphs and statistical tests presented in Section 5 and 6. A summary of the best model for each station and each forecast time lag given the available data is presented at Table 10 (Section 7, p. 57). The numerical developments have been realized in the programming environment MATLAB6.1®. The available data has been transformed into Excel format before using it as input data for the model development and application. The developed software has a graphical user interface. A CD-ROM containing the original software code and a compiled version is attached to the present report. The original source code can only be used in connection with Matlab®. iii The analysis underlying the present report can easily be transposed to other stations of the Magdalena river or include other available data. It can also be applied to other Colombian rivers having the same characteristics. A technician who is used to deal with hydrological forecasts can complete the necessary work and model testing. A specific training will be given to potential users during the visit of the expert accompanied by one of his co-workers. A generalization of the developed models will ask for an extension of the programs in order to enable their automatic use under any circumstances. The availability of the latest Matlab® version will be necessary. These questions will be discussed during the next expert mission. Finally, it is noteworthy that the development of an operational forecasting tool will require an important amount of additional work to integrate the real-time data measured in the field. This data will namely enable a real-time update of the model and its forecasts according to the errors between the observed data and the realized forecasts. Recursive parameter estimation could also highly improve the forecast quality but would need more sophisticated models and accordingly an appropriate training of the users.