Thoughts on “a 6-hour tornado forecast”

Recently, an individual who works for one of the leading U.S. weather forecasting companies made a bold claim, that his company, using a weather forecast model developed at a separate company they recently acquired, produced a forecast of a tornado with 6+ hours of lead-time. While this single forecast certainly had some positive aspects, the notion that a tornado was successfully predicted, and that anyone in the weather enterprise is capable of skillfully predicting such events on time-scales of multiple hours is just not true. Here’s why:

1. In nature, a rotating thunderstorm updraft (i.e. mesocyclone) is not a sufficient condition to produce a tornado. Many thunderstorms that possess mesocyclones do not produce tornadoes, with some estimates suggesting only 25% of mesocyclones produce tornadoes. Anyone who’s chased a supercell thunderstorm or observed one using radar knows this, and it’s part of the reason National Weather Service tornado warnings possess false-alarm rates of 70–80%.
2. The field used to make the forecast, mid-level (2–5 km) hourly-maximum updraft helicity (UH), was designed to identify mid-level mesocyclones in high-resolution model output (using the procedure outlined in Kain et al. 2008, section 2e). Following naturally from point 1, the presence of a mid-level mesocyclone in forecast output should not imply the presence of a tornado in the forecast since this isn’t observed in nature. Large UH magnitudes simply indicate the presence of a rotating thunderstorm in forecast model output, not the presence of tornadoes.
3. The highest-resolution forecast models presently run in operational settings, including those used to produce the “6-hour tornado prediction”, are run at grid spacings (~ 3-km) that barely resolve the mesocyclone and do not resolve finer-scale processes important for tornadogenesis. Estimates of the grid spacing necessary to realistically simulate thunderstorms range from 1-km to as small as 100-m. And even finer resolution is needed to capture tornado-scale details.
4. Uncertainty in timing and placement of thunderstorms are common in convective-scale forecasts, even as short as a few hours. Studies that have shown that UH possesses skill as a predictor of severe thunderstorm activity (e.g. here, here, here) have used verification techniques that allow for spatial and temporal forecast uncertainty. One-to-one correspondence between forecasted thunderstorms and observed thunderstorms is not expected! Although, research efforts are underway to make very accurate short-term (1–2 hr) forecasts of convection where such correspondence is desired (see below).
5. Even if the scientific constraints described above did not exist, suggesting you are capable of 6-hour lead time tornado predictions by cherry-picking one case is grounds for rejection from any reputable scientific journal. Start by running the forecast every day for a season, without any predefined selection of cases, and verify those results.

On the other hand, there are scientists in academia and government labs who recognize these fundamental scientific challenges, and are working to develop methods that could lead to improvements in short-term forecasts and warnings of convective hazards such as tornadoes. Realistic goals within these projects anticipate extending tornado lead-times to one-hour within the next decade, much shorter than the 6-hour lead-time suggested in the aforementioned blog post.

These efforts recognize the need for state-of-the-art data assimilation techniques to produce analyses of thunderstorms and ensemble forecasts to quantify the uncertainty of the forecast evolution. Even the forecast models used in these cutting-edge techniques are too coarse to resolve tornado-scale processes, and it may be years before computing infrastructure can meet these demands (see point 3 above). The 6-hour tornado forecast contained no representation of the forecast uncertainty that is required on the scale of individual thunderstorms. We’ve learned over and over again the problems with disregarding uncertainty information.

In light of the scientific points outlined above, the suggestion of a forecast system producing 6-hour tornado forecasts with any significant frequency is at best misleading, and at worst could be construed as false advertising. This kind of salesmanship could potentially devalue the ongoing, life-saving, work scientists are doing now to improve tornado warnings and forecasts. Selling the abilities of such a product to the public and potential clients is in my opinion, bad business, and I hope no one will be buying.