Not quite sure of the right location for this.
A few people in the various severe weather threads have mentioned
a paper whose lead author is our favourite pink-pencil wielding forecaster (if anyone here doesn't know where the Broyles meme comes from, it's his long standing combo of generally bullish forecasts and aggressive wording). Aims of the project are to improve outbreak forecasting and come up with a consistent way to determine risk categories.
Part of the paper is spent on something called a 'risk impact value', which is an attempt at "estimating the negative community impact of a tornado outbreak" based on track data. This is to provide a more objective way to determine what events 'deserved' a high risk. It includes weightings for number, length, strength etc, with results evaluated by other forecasters. Unfortunately, although they give some details (e.g. that each criterion is based on percentiles, and that they're weighted), they don't give the actual formula/e, or full details of how they constructed the model.
There's some GIGO-potential. For example, using seven F5s for the Super Outbreak (it was six, according to the person who actually assigned the ratings) and the obvious inconsistencies of ratings over the offical dataset, undercounting, and the recording of families as continuous paths (particularly as consideration was given for events over 100 miles). Nonetheless, the initial version produces an agreeable ordering of the events they tested it on, as shown in their figure 2 and 'RIV Top 20 Events (2000 to 2024)' table.
Where things are a bit odd though, is when it is applied to the full dataset (apologies for the wonky columns):
Top 20 Events (1950 to 2024)
Rank Date RIV Deaths
1: April 3, 1974 405.8 310
2: April 27, 2011 252.8 316
3: May 31, 1985 169.4 76
4: May 27, 1973 136.2 9
5: April 11, 1965 125.6 260
6: June 7, 1984 116.6 13
7: June 2, 1990 116.1 9
8: March 13, 1990 106.1 2
9: June 4, 1955 97.9 0
10: April 12, 2020 93.2 30
11: March 21, 1952 87.5 205
12: June 8, 1974 87.5 22
13: March 31, 2023 85.4 23
14: April 2, 1982 85.3 30
15: May 24, 2011 82.0 18
16: June 16, 1992 80.8 1
17: April 26, 1991 73.7 21
18: February 5, 2008 69.7 57
19: Nov. 22, 1992 68.2 10
20: May 20, 1957 68.0 44
Most of that's good and fine, albeit the placing of 31/5/85 over Palm Sunday is a surprise (and makes me think F/EF5 has a heavy weighting). But there's a couple of oddities.
The first is the 27/5/73 outbreak, ranking ahead of more famous outbreaks like Palm Sunday 1965. It produced, officially, 24 significant tornadoes, which is high but not exceptional (many on the list had more). Many of these have no path length and Grazulis only lists fifteen. All three F3+ long track tornadoes were probably families. I'd guess there's a heavy weight on the 135 mile Brent EF4. Grazulis lists this is a family, but one witness claims they followed it continuously in a plane. Ordinarily that would be good enough for me, but LandSat agrees with Grazulis. I suspect better data would knock this one off the list.
The second is June 4 1955, which is ranked ninth. I don't understand how this made the list with six significant tornadoes. There's two long track F4s, though Grazulis (almost certainly correctly) gives them much shorter paths. Is it a mistake? Is the coverage criterion throwing things off?
Lastly, where's 10/12/21?
In all an interesting section when it comes to ranking and rating major outbeaks, but the lack of the actual algorithm and a couple of odd details make me hope there's further articles that explain things better in the future.
www.leelanauhistory.org
