This paper investigates if the use of neural networks can improve the accuracy of tree-ring based paleoclimatic reconstructions with respect to some of the commonly used methods. A three layers feedforward model of neural network is shown to be very efficient in explaining a high percentage of the variance of the instrumental climatic record both for calibration and validation. Some traditional statistics have been estimated to evaluate the accuracy of the reconstruction. These results have been finally compared with those of a regression-based model, showing the higher accuracy of the neural network reconstruction. ¿ 2000 American Geophysical Union