Making tropical cyclone intensity forecasts more reliable

Guest author: Xin Guan

When a tropical cyclone is bearing down on a region, there are a whole range of actors responsible for managing different parts of the risk chain. Re/insurers are focused on understanding portfolio exposure, anticipating losses and preparing to respond to claims, whereas emergency managers are working before and during the event to protect people, infrastructure and critical services on the ground. Humanitarian agencies are focused on anticipating where needs will be greatest and directing resources towards the communities most at risk. All parties have equally urgent and critical roles to play, and they all need forecasts they can actually act on. Crucially, those forecasts also need to align at the handoffs between different stages of the risk chain, so that decisions made by one actor do not become disconnected from those made by another.

That’s the goal behind PsiClone, a disaster response action management platform we released earlier this summer. At its core, the tool is designed to help different sectors respond to hazards quickly and effectively. Currently focused on tropical cyclones, the tool ingests live ECMWF ensemble forecasts to produce a continuous ensemble forecast that has visibility up to 15 days ahead of an event.

As part of this work, PhD student Xin Guan joined Maximum Information for 6 months as part of an ongoing partnership with the University of Bristol. Her work on this product focused on solving an important problem that arises when producing a feature like this: raw forecasts from Numerical Weather prediction (NWP) models tend to be systematically biased. This means that, without correction, the forecast can consistently over- or under-estimate aspects of a storm’s behaviour, intensity or timing. Correcting these biases is therefore important if the forecast is going to provide a reliable basis for decisions made ahead of and during an event.

Below, Xin provides an overview of her work; focusing on the bias correction on wind intensity, what methods she looked at, the choices that were made and the challenges that users still need to be aware of when handling data that has been bias corrected.

Why raw forecasts fall short?

For tropical cyclone intensity specifically, global models consistently under-estimate peak wind speeds. The main reason for this bias is the coarse spatial resolution of the models, which means that variables, including ground level wind speeds, are averaged over areas spanning several kilometres. As a result, the model grids are therefore often too coarse to capture the small, violent core of a storm. In addition, the physical processes governing interactions between the atmosphere and ocean are necessarily approximated, and the models are initialised from an imperfect snapshot of these complex systems. The result is that the strongest winds of a storm are underestimated. This is far from ideal for anyone pricing or managing catastrophe risk.

The fix: statistical correction on forecasts

Rather than rebuilding weather models, the usual practice is to correct their output after it’s produced by statistical post-processing. We learn from the gap between past forecasts and what actually happened, then apply that lesson to new forecasts. Throughout our work on this project, our team compared two classes of methods:

Quantile-based methods, which reshape the whole distribution of forecasts to better match the distribution of observations.
Regression-based methods, which model the relationship between forecast and observation directly.

This was tested across several ocean basins on the Northern Hemisphere (the North Atlantic, the Northwest and Northeast Pacific, and the North Indian), since the storms’ behaviour, and model’s errors vary from region to region.

The problems we ran into, and how we solved them

1. Regular model updates.
Weather centres regularly upgrade their models, so error patterns from years ago may no longer apply. We addressed this with a rolling six-year training window, so corrections always reflect the recent behaviour of the current system. It also shaped our method choice: standard Quantile Mapping assumes past and present distributions line up, which breaks when the model keeps changing, so we switched to Quantile Delta Mapping (QDM), which learns from model updates.

2. Errors depend on how far ahead you’re forecasting.
We handled this by letting the correction vary smoothly with lead time, adding a smooth function on top of both QDM and the regression-based method.

3. Simple relationships weren’t flexible enough.
We moved from Simple Linear Regression (SLR) to a Generalised Additive Model (GAM), which captures more flexible patterns. We also found that adding latitude as a factor meaningfully improved the corrections, since storm behaviour genuinely differs by location.

4. The most extreme storms are rare.
Because very high wind speeds show up so rarely in the historical record, the standard corrections still underserve the catastrophic cases that matter most. To push further, we built a peak-weighted GAM that pays extra attention to the high-intensity end.

What we chose, and the trade-offs

Three models have been implemented in the PsiClone platform, each with a different strength:

Smoothed QDM: produces a wider spread, so for extreme events it’s more likely to have ensemble members that exceed the observed peak.
GAM with latitude: has the best performance in terms of Mean Absolute Error (MAE), however the spread is not increased, and it still under-predicts intense storms.
Weighted GAM with latitude: reduces the under-prediction of intense winds the most, at the cost of over-correcting weaker storms.

Plots showing raw forecasts and bias corrected forecasts for hurricane Ian, forecast issued five days before storm peak. Each panel applies a different bias-correction method to the same ensemble forecast (cyan), compared against observations (red); the "Raw forecasts" panel shows the uncorrected baseline. The weighted GAM lat corrected forecasts get closest to the observed peak for a severe storm as hurricane Ian, which peaked as a Category 5 hurricane.

Plots showing raw forecasts and bias corrected forecasts for hurricane Fred, forecast issued five days before storm peak. Each panel applies a different bias-correction method to the same ensemble forecast (cyan), compared against observations (red); the "Raw forecasts" panel shows the uncorrected baseline. Here the weighted GAM over-corrects for a lower-intensity storm like Fred, which peaked as a Category 1 hurricane, while GAM lat performs best, tracking the observed peak most closely with a reasonable spread.

User Guidance

There’s no single winner, the right correction model is task-specific.

You’re worried about being blindsided by the worst case.
Use smoothed QDM. It produces a wider range of outcomes, so its ensemble is more likely to include members that reach or exceed a storm’s true peak. This is the choice when you want the forecast distribution to represent plausible worst-case scenario.
You want the most accurate single estimate of a storm’s likely intensity.
Use GAM with latitude. It has the lowest average MAE, so it’s the most appropriate choice for day-to-day monitoring and central-estimate planning.
You’re specifically focused on the most intense storms.
Use weighted GAM with latitude. It corrects the under-prediction of severe winds most aggressively, making it the best fit when the systematic downward bias in catastrophic, high-intensity events is your main concern. The trade-off is that it tends to over-correct weaker storms, so it’s less suitable as a general-purpose tool.

Problems that didn’t get solved by bias correction

If raw forecasts didn’t predict a peak, bias corrected forecasts won’t produce a peak either.
If there is a delay in the peak forecasted, which it’s often the case, bias correction can-not address it, although GAM latitude seems to improve it slightly.
For weak storms, which are usually predicted well by raw forecasts, any bias correction will lead to over-correction.

Contact Us