Abstract: Although probability forecasts based on a parametric regression scheme have good fitting rates the results are not so stable. For this reason, a new approach is proposed to such forecasts by means of a K-nearest neighbor nonparametric regression technique, and the technique includes 4 main components such as a database of historical samples, production of nearest neighbor subsets, their optimization and estimate of predictands. Case experiments are conducted on univariate (cloudiness or precipitation) and multivariate joint (e. g., rainfall, total cloudiness, wind speed and temperature) probability forecasting, with the results tested. Results show that forecasts from the nonparametric regression scheme are high-stability, with good prospects in operational weather forecast.