The second implication of the CLP model is that there is a relationship between landfalling hurricane numbers and the number of hurricanes in the basin. Given this, it might make sense to predict landfalling hurricane numbers by first predicting basin hurricane numbers, and then converting the basin prediction to a prediction of landfalling numbers using the probability of landfall estimated from historical data since 1950. We will apply this idea to all subsequent models (thus doubling the number of predictions) and will call such predictions 'indirect' predictions, as opposed to 'direct' predictions which predict landfalls directly from landfall data. Should we expect such indirect predictions to be better than direct predictions? In Laepple et al. (2007b) we find that, if the probability of storms making landfall is constant, then there are situations in which we would expect the indirect methods to be better. For instance, if we make a short baseline prediction of the number of basin storms using data since 1995, and then convert that to a prediction of land-falling storms using a probability of landfall estimated from the number of basin and landfalling storm numbers since 1950, theoretically we expect the resulting prediction to be nearly twice as accurate as the direct short baseline prediction that uses landfalling data alone. The explanation for why indirect methods beat direct methods in this case is that there are more storms in the basin, and so the predictable signal can be estimated more accurately than when only using the landfall data. If the probability of landfall can also be estimated relatively accurately (as it can if we use data since 1950), then this increased accuracy propagates through to the prediction at landfall. In practice, our backtesting studies show that the indirect predictions do lead to better predictions, but are not usually twice as accurate as the direct predictions. Model skill is discussed in section 8 and Fig. 4 shows a comparison of model errors. Row 6 in Tables 2 and 3 shows the indirect short baseline predictions.
Was this article helpful?