Logo Wind-focused Hurricane Interactive Track Simulator

Licensing by Columbia Technology Ventures

An accurate and unbiased global 10,000-year stochastic catalog of tropical cyclone (TC) tracks and wind data generated by the Wind-focused Hurricane Interactive Track Simulator (WHITS) is now available for download. WHITS is specifically developed to provide reliable and unbiased global TC risk assessments over an operational time frame.

The model preserves storms' historical memory and produces realistic tracks, track densities, and winds. WHITS improves on its predecessor HITS with additional wind parameters and smoother transitions to better support risk applications. The download package includes the full stochastic catalog, all supporting Python code used for data generation, and the figures presented below. Additionally, customized risk assessments and modeling solutions are available through sponsored research opportunities. Click here to explore our current sponsored research initiatives.

Please contact Dovina Qu, techtransfer@columbia.edu for licensing of the stochastic catalog package or to discuss sponsored research tailored to your needs.

Unbiased Tracks and Landfalls

Limited historical data on TC landfalls makes it difficult to accurately estimate and correct model biases, especially for rare, high-impact events. Simple bias corrections often assume uniformity across regions and storm intensities, which is rarely accurate, and correcting for specific locations is especially challenging due to data sparsity and uncertainty. Overfitting is a risk, and ignoring dependencies between landfall frequency and storm characteristics can produce unrealistic results. The WHITS model addresses these issues by using machine learning to adapt and reassemble historical storm track segments, preserving track memory while generating 10,000 realistic synthetic seasons for global insurance risk analysis and coastal planning while avoiding bias correction issues. The resulting tracks are not exact reproductions of historical segments, they are adapted from historical data to improve realism and applicability for risk modeling.

Performance

STORM stochastic catalog of tropical cyclone tracks, including five season tracks plots, track density plots, and wind probabilities plots here to compare directly to WHITS below.

Realistic simulated storm tracks and wind speeds for each ocean basin. These maps display the first five simulated tropical cyclone seasons. Each track begins at an open circle (genesis point) and is represented by a line colored according to wind speed.

Observed track density (left) and simulated track density (right). Simulated is the 50th percentile (median) of 100 random draws of the same number of seasons as the observed. The sample median is a strong choice because it is not affected by extreme values or outliers, making it a robust measure of central tendency even when the underlying distribution is skewed or has extreme values.

Observed Annual Maximum Wind >=64 knots Probability (left) and simulated (right). Simulated is the 50th percentile (median) of 100 random draws of the same number of seasons as the observed.

Further validation for the prior HITS version is provided here:
Plots of 6-hour intervals per year (see Figure 7)
Landfall frequency (see Figure 8)
Wind speed distributions (see Figure 9)
Observed hurricanes (see Figures 10 and 11)
These figures together showcase important metrics like temporal distribution, landfall frequency, wind speed patterns, and comparisons to real hurricane events.

Methodology

WHITS uses the 3-hourly IBTrACS dataset and retains the original timestamp and IBTrACS identifier for each observation on every segment.

HITS and WHITS both use the bisquare kernel function (equation 5b) to calculate the probability of jumping to another track, considering distance (5c), comparative vector (5d), and age (5e). WHITS adds wind speed as an additional variable. The outer exponent, or degree, of the bisquare kernel for each variable determines how unlikely it is to make a jump to a very different track.

In WHITS, the kernel degree is increased from 2 to 6 for both the comparative vector and wind, while distance and age remain at degree 2. For distance and age, the probability at the outer edge (beyond the 90th percentile) is 0.0361. For the comparative vector and wind, after testing various exponents and settling on 6, the probability at the outer edge drops to 0.000047. This means that while such jumps are not impossible, they are extremely unlikely—about three orders of magnitude less likely to jump to a track with a large wind difference compared to using a degree of 2.

The difference between the point to jump to and the next point on the original track is computed as delta_lon and delta_lat. Segments are shifted by delta_lon and delta_lat to create more continuous tracks. Jumps are smoothed further in a window 6 hours before and after. Note that it is important that the smoothing does not occur over the whole simulated track.
  • 90° turns - examples N. Atlantic: Charley (2004), Wilma (2005), Babe (1977), Elena (1985), Michael (2018), Franklin (2023), W. Pacific: In-fa (2021), Lionrock (2016 - both 90° turn and loop), Jongdari (2018), C. Pacific: Walaka (2018)
  • hair pin turns - examples N. Atlantic: Ingrid (2013), Joaquin (2015), Alice (1954/55), Idalia (2023), W. Pacific: Noru (2017)
  • loop de loops - examples N. Atlantic: Jeanne (2004), Betsy (1965), Danielle (2022 - 2), Don (2023), Nigel (2023), Katia (2023), Margot (2023), Rafael (2024), W. Pacific: Wayne (1986), Opal (1962), Vera (1959)
  • and other unique shapes - examples N. Atlantic: a pretzel Gert (2023), a bow Juan (1985), a figure eight Leslie (2018), a leaf Tropical Storm Karl (2022)
In the segments between jumps, are retained. We have an unbiased sampling of natural track behavior.

Storm tracks during the 2023 North Atlantic tropical cyclone season (below from NASA/NHC) showcase the diverse patterns that occur naturally.

Who we are

  • Jennifer Nakamura
    • Research Scientist in Ocean and Climate Physics at the Lamont-Doherty Earth Observatory of Columbia University. Her background is in atmospheric science and hydroclimatology. She studies the extremes of precipitation in both floods and droughts and their respective causes (like TC!), along with climate variability and ocean-atmosphere interactions.
  • Upmanu Lall
    • Visiting Professor at the Columbia University Department of Earth and Environmental Engineering and founder of the Columbia Water Center. Professor at the School of Complex Adaptive Systems and Director of the Water Institute at Arizona State University. He has broad interests in hydrology, climate dynamics, water resource systems analysis, risk management and sustainability, machine learning, deep learning, and AI. He is motivated by challenging questions at the intersection of these fields, especially where they have relevance to societal outcomes or to the advancement of science toward innovative application.