Abstract Simple and multiple discriminant models of the first stage episode day for the 30-, 24- and 6-h prediction intervals have been derived for Azusa, Burbank, Los Angeles, Fontana and San Bernardino, and those of the second stage episode day have been developed for Azusa and Fontana. Morning weather variables and the previous day's maximum 1-h average oxidant concentration are used as discriminating variables. A scattergram, Pearson and canonical correlation coefficients, and Wilks' Lambda are presented to show the statistical relationship between weather variables and oxidant concentrations or episode occurrences. The morning sounding profiles taken at Los Angeles International Airport differ markedly between episode and nonepisode days at various air monitoring stations. The previous day's oxidant concentration, temperatures at 950- and 850-mb. height of inversion base, inversion magnitude and inversion breaking temperature correlate significantly high with the oxidant concentration at all stations and are important discriminating variables for predicting the occurrence of episode and nonepisode days. It is found that the models yield approximately 65–88% accuracy for the first stage episode day and 51–80% accuracy tor the second stage episode day. In the most polluted area, the multiple discriminant model provides very little incremental prediction power over the simple model using the 850-mb temperature or 24-h persistence variable as a predictor for the first stage episode day. but it provides a larger incremental prediction power for the second stage episode day.