We present a methodology for estimating on-line the cell-loss probability of an ATM link. It is particularly suitable for estimating small probabilities of the order $10^{-6}$--$10^{-9}$, with variance many orders of magnitude smaller than traditional estimators. The method is justified by the theory of large deviations, and the information required is based upon the actual traffic flows rather than upon the analysis of some specific traffic model. The method is effective when there is a large degree of statistical multiplexing; in other words, when the number of input traffic sources to be large. The statistical properties we require for the traffic are very general and are met by most real-time traffic source models. Experimental results suggest that the estimators obtained are extremely timely and accurate in the case of large systems, and hence that the method can be used for the on-line estimation and control of the congestion in the actual network.