The bulk of existing work on the statistical forecasting of air quality is based on either neural networks or linear regressions, which are both subject to important drawbacks.