Mixed Baseline Predictions

Up to this point, we have presented averaged baseline prediction methods. The long and medium baseline prediction methods have the advantage that they use a large amount of historical data, but since, in the CLP model, we believe that the land-falling hurricane number time-series is non-stationary, we expect that the longbaseline and medium-baseline predictions are biased. The short-baseline scheme only uses the most recent data, which is likely to be more relevant to the current climate, and so is likely to reduce the bias, but possibly suffers from the fact that there are simply not many years of data since the last change-point to make an accurate estimate of the mean number of storms. Consideration of the pros and cons of the long and short baseline models motivates the idea that there may be a forecast methodology that lies 'in between' the two, and performs better than both. We have

Historical MDR SST (C) and 5-Year SST Model Forecasts

Historical MDR SST (C) and 5-Year SST Model Forecasts

1975 1980 1985 1990 1995 2000 2005 2010 Year

Fig. 4 a) The SSTs in the MDR region since 1880. b) Predictions of the 2006-2010 MDR SSTs using flat-line (diamond), damped trend (middle prediction) and linear trend (triangle) predictions

1975 1980 1985 1990 1995 2000 2005 2010 Year

Fig. 4 a) The SSTs in the MDR region since 1880. b) Predictions of the 2006-2010 MDR SSTs using flat-line (diamond), damped trend (middle prediction) and linear trend (triangle) predictions investigated this idea in a series of technical reports [Jewson et al., 2005; Jewson et al., 2006 and Binter et al., 2006]. The methodology we use is to formulate the question as a classical mathematical 'bias-variance trade-off' problem, and ask what weights we should put on the different parts of the historical hurricane data in order to minimize the RMSE of our predictions. We call predictions from the resulting method 'mixed-baseline predictions'. They are an example of a statistical modelling methodology known as 'shrinkage'. As with the short baseline predictions, we have two versions of each mixed baseline prediction: one direct (determined by applying weights directly to the historical landfall number data) and one indirect (determined by applying weights to the historical basin number data, and then converting the basin prediction to a landfall prediction using an estimated proportion).

To estimate the optimal weights on the mean levels of activity between change-points we use the weights which minimize the RMSE of hurricane number predictions (Binter et al., 2006). We find that the lowest RMSE prediction for the current level of category 3-5 storms comes from a set of weights corresponding to a straight average of the two most active periods in the historical record for both the Elsner and the RMS change-points. For the Elsner change-points these are 1943-1964 and 1995-2005 and for the RMSchange-points these are 1932-1947 and 1995-2005. The mixed baseline model analysis tells us that this gives a more accurate prediction than just using data from the recent active period because it makes use of more historical data, while only introducing a small bias. Interestingly, for the category 1-5 basin storms, the best prediction comes from a set of weights corresponding to just using the historical data since the last change point, because there are enough cat 1-5 basin storms since 1995 that it is no longer beneficial to try and reduce the variance of the forecast using earlier data.

In order to address the question of how sensitive these predictions are to the change-points identified in the historical hurricane number time-series, we apply mixed baseline methods to the change-points from both Elsner et al. (2004) and from Jewson and Penzer (2006). The four combinations (two sets of change-points and both direct and indirect methods) lead to the predictions given in rows 4, 5, 7 and 8 of summary tables, 2 and 3 Note that the indirect predictions of the category 1-5 hurricane numbers for the Elsner and RMS change-points will be the same as the indirect predictions of the short baseline since the optimal mixed baseline for basin category 1-5 hurricane numbers is just the short baseline.

Was this article helpful?

0 0
Survival Treasure

Survival Treasure

This is a collection of 3 guides all about survival. Within this collection you find the following titles: Outdoor Survival Skills, Survival Basics and The Wilderness Survival Guide.

Get My Free Ebook


Post a comment