[Paper Review] Year-ahead prediction of US landfalling hurricane numbers: intense hurricanes

Shree Khare, Stephen Jewson|ArXiv.org|Dec 10, 2005

Tropical and Extratropical Cyclones Research3 references3 citations

TL;DR

This study develops a year-ahead statistical prediction model for intense U.S. landfalling hurricanes using historical landfall counts, with a focus on optimizing the length of time-averaging windows. It finds that longer averaging windows (16–56 years) are optimal for predicting intense hurricanes—contrasting with shorter windows (6–33 years) for all hurricanes—suggesting distinct predictability dynamics due to lower signal-to-noise ratios in intense storm data.

ABSTRACT

We continue with our program to derive simple practical methods that can be used to predict the number of US landfalling hurricanes a year in advance. We repeat an earlier study, but for a slightly different definition landfalling hurricanes, and for intense hurricanes only. We find that the averaging lengths needed for optimal predictions of numbers of intense hurricanes are longer than those needed for optimal predictions of numbers of hurricanes of all strengths.

Motivation & Objective

• To improve year-ahead prediction of intense U.S. landfalling hurricanes using a more accurate definition of landfalling storms.
• To assess whether the predictability characteristics of intense hurricanes differ from those of all hurricanes.
• To identify optimal averaging window lengths for forecasting intense hurricane numbers using historical data.
• To evaluate the robustness of results across different data periods and statistical tests.
• To inform practical forecasting by reducing arbitrariness in window length selection through backtesting and likelihood-based weighting.

Proposed method

• Uses the SSS variable from the HURDAT database to define a landfalling hurricane as one that is a hurricane at the moment of landfall, improving accuracy over prior definitions.
• Applies a backtesting framework to evaluate time-averaging windows of varying lengths (from 6 to 100 years) using historical data from 1940–2004 and 1900–2004.
• Measures prediction skill using mean squared error (MSE), identifying the window length that minimizes MSE as the optimal predictor.
• Performs statistical significance tests via random reordering of the hurricane count time series to validate the optimality of selected window lengths.
• Decomposes MSE into bias and variance components, further splitting variance into internal variability and sampling error to diagnose error sources.
• Conducts sensitivity analyses using 41 overlapping data series starting from 1900 to 1940, all ending in 2004, to test consistency of results.

Experimental results

Research questions

• RQ1 What is the optimal time-averaging window length for predicting the number of intense U.S. landfalling hurricanes one year in advance?
• RQ2 How do the predictability properties of intense hurricanes differ from those of all hurricanes in terms of optimal averaging window length?
• RQ3 Does the use of a revised definition of landfalling hurricanes (SSS vs. XING) significantly alter the prediction performance or optimal window length?
• RQ4 What proportion of forecast error in intense hurricane predictions is attributable to sampling error versus internal variability?
• RQ5 Can the wide range of optimal window lengths be reduced through likelihood-based weighting of forecasts to improve practical forecast reliability?

Key findings

• For intense U.S. landfalling hurricanes, the optimal averaging window length ranges from 16 to 56 years, significantly longer than the 6–33 year range found for all hurricanes.
• The minimum mean squared error (MSE) for intense hurricanes occurs at a 16-year window, with a p-value of 3.35% from the random reordering test, indicating statistical significance.
• Short averaging windows (e.g., <10 years) produce high forecast errors primarily due to sampling error, which dominates the variance component of MSE.
• The broader optimal window range for intense hurricanes suggests a lower signal-to-noise ratio compared to all hurricanes, potentially due to their lower frequency.
• Results are robust across multiple data periods (1940–2004, 1900–2004, and 41 overlapping 1900–1941 series), confirming consistency in the optimal window range.
• The study identifies a need for likelihood-based weighting of forecasts across window lengths to reduce arbitrariness in selecting a single optimal window for practical use.