[Paper Review] An objective change point analysis of landfalling historical Atlantic hurricane numbers

Stephen Jewson, Jeremy Penzer|ArXiv.org|Nov 8, 2006

Tropical and Extratropical Cyclones Research5 references4 citations

TL;DR

This study applies an objective change point analysis to U.S. landfalling Atlantic hurricane counts using a brute-force optimization of out-of-sample mean squared error (MSE) to detect structural shifts. Despite physical plausibility of climate-driven variability, the analysis finds no statistically significant change points in landfalling hurricane frequency, with results inconsistent and unstable compared to basin-wide hurricane data, suggesting insufficient signal-to-noise ratio in the landfalling record for reliable detection.

ABSTRACT

In previous work we have analysed the Atlantic basin hurricane number time-series to identify decadal time-scale change points. We now repeat the analysis but for US landfalling hurricanes. The results are very different.

Motivation & Objective

• To detect statistically significant change points in the annual time series of U.S. landfalling Atlantic hurricanes using an objective, optimization-based method.
• To assess whether the landfalling hurricane record contains detectable temporal shifts comparable to those found in basin-wide hurricane counts.
• To evaluate the stability and reliability of detected change points in a low-count, noisy time series with many zero years.
• To compare results with prior studies using different methodologies, particularly Elsner et al. (2004) using Markov Chain Monte Carlo.
• To explore whether intense landfalling hurricanes show more detectable change points than total landfalling counts.

Proposed method

• Employs a brute-force search over all possible combinations of change points to minimize an out-of-sample mean squared error (MSE) cost function.
• Imposes a minimum 10-year gap between change points to make the combinatorial search computationally feasible.
• Uses the HURDAT database to construct a time series of hurricanes that are at least Category 1 at landfall, excluding weakening storms.
• Applies statistical significance testing via 100 random reorderings of the data to assess whether the observed minimum MSE is unlikely under the null hypothesis of no change points.
• Compares results across models with 1 to 4 change points, evaluating stability through top 30 solutions per model.
• Extends analysis to intense landfalling hurricanes (Saffir-Simpson category 3 or higher) to test sensitivity to intensity thresholds.

Experimental results

Research questions

• RQ1 Are there statistically significant change points in the annual time series of U.S. landfalling Atlantic hurricanes?
• RQ2 How do the detected change points in landfalling hurricane counts compare to those in the total basin hurricane count?
• RQ3 Is the signal of change points more detectable in intense landfalling hurricanes than in total landfalling counts?
• RQ4 How stable are the detected change points across different model configurations and resampling methods?
• RQ5 To what extent do the results align with prior studies using alternative statistical methods, such as Markov Chain Monte Carlo?

Key findings

• No statistically significant change points were detected in the U.S. landfalling hurricane count time series, as the minimum out-of-sample RMSE did not show a meaningful improvement over a constant-rate model.
• The best two-level model performed only slightly better than the one-level (constant-rate) model, and the improvement for models with more than two change points was minimal and not statistically significant.
• The optimal change points identified—such as a reduction in 1956 or shifts in the 1970s and 1980s—did not align closely with those found in basin-wide hurricane data, indicating weak correspondence.
• Change point estimates were unstable across the top 30 solutions, particularly compared to the more stable results from basin-wide hurricane analysis.
• For intense landfalling hurricanes, the RMSE score decreased slightly with more change points, but the minimum was not statistically significant, as the real data’s best score (0.840) was worse than the mean of 100 random reorderings (0.822).
• The results are consistent with Elsner and Jagger (2004), who also found no strong evidence for change points using a different algorithm, reinforcing the conclusion that the landfalling record lacks sufficient signal for reliable detection of structural shifts.