MDA was compared to conditional climatology (CC), persistence and pure climatology. Preliminary results indicated that ceiling and visibility forecasts from MDA and CC were equivalent. These forecasts were largely superior to the other two techniques.
Several avenues to improve the MDA forecasts were considered. First, the predictors were all categorised, as in CC, to remove some unnecessary variability. Second, a multi-step procedure was designed in which a first MDA is trained to separate the highest (more frequent) category from the others. Then, another MDA was developed using only a data-base without observations in the highest category. A conditional probability was applied to the results from that second MDA to produce a final forecast. This was done because the database is dominated by observations in the highest category. As a result, statistical forecasts are attracted toward the highest category, particularly for longer projection times. A unit-bias procedure, in which thresholds are set to forecast each category with the same frequency at which it is observed, was considered. Third, experiments were conducted to determine the number of predictand categories that the available dataset could support. Originally, 9 ceiling and 7 visibility categories were needed for operations, but it proved difficult to discriminate among this many categories at some stations.