Atlantic hurricane forecast: a statistical analysis

Purpose – A reliable forecast of hurricane activity in the Atlantic Basin has the potential to help mitigate the economic losses caused by hurricanes. One of the difficult problems is to make reasonable annual forecast of catastrophe losses based on the short record of historical observations. Atmospheric conditions tend to influence tropical cyclone development. Considering the complex interactions among climatological factors, prediction of future hurricane activity is challenging. In this study, the authors are attempting to predict the number of Atlantic hurricanes for a given year based on two different approaches. Design/methodology/approach – In part I, an autoregressive integrated moving average (ARIMA) is used to model a long‐run behavior of Atlantic hurricane frequency. The authors present a comparison of CSU's forecast with ARIMA model. Part II focuses on the relationship between the climate signals and hurricane activity and introduces a new approach in including climate indices into the prediction model. In this part, principal components analysis (PCA) is used to identify possible patterns in historical data based on six climate indices measured prior to hurricane season. The objective is to reduce the data set to a smaller set while most of the variability observed in the real data is captured. The variances observed in an orthogonal system indicate the order of contribution of each mode shape. Findings – Results from part I suggest that CSU's forecast model, in general, is superior to results obtained by ARIMA. In part II, the correlation between mode (shapes) and the number of Atlantic hurricanes per year is examined. The resulting relationships show that, for the time interval of 1990 through 2011, PCA‐based approach provides better estimates compared to CSU's forecast. Originality/value – The paper presents a unique prediction approach which is simple, relatively accurate and easy to apply. The results of this study show that complex statistical analyses/models do not necessarily provide better forecasts.