James Franklin
@franklinjamesl.bsky.social
Former Chief, Hurricane Specialist Unit, National Hurricane Center, NOAA/NWS (retired). Mostly weather posts, with occasional forays into grammar and usage, word play, and south Florida sports.
Here's a general description of the interpolator from a recent manuscript. I don't think NHC's actual taper parameters are documented anywhere but I don't think they're exactly secret either. Let me see if there's any objection to sharing them.
November 6, 2025 at 6:29 PM
Here's a general description of the interpolator from a recent manuscript. I don't think NHC's actual taper parameters are documented anywhere but I don't think they're exactly secret either. Let me see if there's any objection to sharing them.
Here are updated verifications for 2025, AL/EP combined, EMXI not available. Impressed that NHC is beating everything, including consensus, for intensity, and everything but Google DeepMind (GDMI) for track. Rough year for GFS and the statistical intensity aids. Banner year for GDMI.
October 31, 2025 at 4:52 PM
Here are updated verifications for 2025, AL/EP combined, EMXI not available. Impressed that NHC is beating everything, including consensus, for intensity, and everything but Google DeepMind (GDMI) for track. Rough year for GFS and the statistical intensity aids. Banner year for GDMI.
The sequence of NHC forecasts verifying at 28/18Z, roughly the time of landfall, is quite remarkable given the spread of the guidance and the track turns involved. Five days worth of OFCLs consistently focused on the western ⅔ of Jamaica, and with remarkable steadiness in timing. Nicely done NHC!
October 30, 2025 at 12:47 AM
The sequence of NHC forecasts verifying at 28/18Z, roughly the time of landfall, is quite remarkable given the spread of the guidance and the track turns involved. Five days worth of OFCLs consistently focused on the western ⅔ of Jamaica, and with remarkable steadiness in timing. Nicely done NHC!
Regarding the sonde from yesterday morning with the 219-kt spot wind, highest dropsonde wind ever seen in a TC, AOML/HRD shared the raw data and I took a look this morning. They’ll have the final call, but I see nothing wrong with the ob and I suspect it’s going to hold up.
October 29, 2025 at 5:54 PM
Regarding the sonde from yesterday morning with the 219-kt spot wind, highest dropsonde wind ever seen in a TC, AOML/HRD shared the raw data and I took a look this morning. They’ll have the final call, but I see nothing wrong with the ob and I suspect it’s going to hold up.
Brad, those 10-m adjustments from the 0-500 m and lowest 150-m layers are based on my analyses of mean dropsonde profiles and have been in place at NHC for the past 20-25 years.
October 27, 2025 at 5:40 PM
Brad, those 10-m adjustments from the 0-500 m and lowest 150-m layers are based on my analyses of mean dropsonde profiles and have been in place at NHC for the past 20-25 years.
Google DeepMind (GDMI) has done real well this year. When NHC today mentioned the GDM ensembles showing most reaching C5 intensity, I was curious how GDMI was doing for RI. Using the HFIP RI metric, turns out only so-so. HMON, HCCA, and OFCL are the standouts for RI so far. Sample still small.
October 25, 2025 at 6:40 PM
Google DeepMind (GDMI) has done real well this year. When NHC today mentioned the GDM ensembles showing most reaching C5 intensity, I was curious how GDMI was doing for RI. Using the HFIP RI metric, turns out only so-so. HMON, HCCA, and OFCL are the standouts for RI so far. Sample still small.
When you do an apples to apples early-only verification, GDMI is still very good, but is not quite as good as OFCL for intensity. (Plus my results are for AL and EP combined for a more robust sample.)
October 22, 2025 at 6:59 PM
When you do an apples to apples early-only verification, GDMI is still very good, but is not quite as good as OFCL for intensity. (Plus my results are for AL and EP combined for a more robust sample.)
The homogeneous sample isn't huge, but this is what is in the public decks for EAII vs GDMI. Also EAII is at a disadvantage since it's a 12-h interpolation half the time while GDMI is (almost) always a 6-h interpolation.
October 13, 2025 at 3:22 PM
The homogeneous sample isn't huge, but this is what is in the public decks for EAII vs GDMI. Also EAII is at a disadvantage since it's a 12-h interpolation half the time while GDMI is (almost) always a 6-h interpolation.
With about 80% of the precincts reporting I've seen enough. GDMI - Google DeepMind is going to win the seat for best track model in 2025. The race for best intensity model is still too close to call, but GDMI is right there with the consensus and OFCL. Quite a remarkable campaign.
October 13, 2025 at 2:45 PM
With about 80% of the precincts reporting I've seen enough. GDMI - Google DeepMind is going to win the seat for best track model in 2025. The race for best intensity model is still too close to call, but GDMI is right there with the consensus and OFCL. Quite a remarkable campaign.
I don't suppose we original season ticket holders are getting another one of these...
October 6, 2025 at 9:46 PM
I don't suppose we original season ticket holders are getting another one of these...
While there's a break in the action I thought I'd update the guidance verification for the season (alas, with EMXI not included). Google DeepMind maintaining its edge with track and competitive with the better models for intensity. HMON also having a fine year.
October 3, 2025 at 11:14 PM
While there's a break in the action I thought I'd update the guidance verification for the season (alas, with EMXI not included). Google DeepMind maintaining its edge with track and competitive with the better models for intensity. HMON also having a fine year.
While we wait for Imelda's shoe to drop, here's how the Google Deepmind (GDMI) has done this year against the regular guidance (apart from EMXI, which isn't in the public decks). For track, GDMI is a bit ahead of everything else, while for intensity it's in the middle of the pack.
September 26, 2025 at 6:42 PM
While we wait for Imelda's shoe to drop, here's how the Google Deepmind (GDMI) has done this year against the regular guidance (apart from EMXI, which isn't in the public decks). For track, GDMI is a bit ahead of everything else, while for intensity it's in the middle of the pack.
Except it needs to make a detour through Alabama first.
September 24, 2025 at 10:09 PM
Except it needs to make a detour through Alabama first.
Ferran, wonder what you think of these FNV3 (GDMI) intensity forecasts for Erin. I think it's fascinating how it got the longer-scale (multi-day) changes practically dead on, but couldn't/didn't catch the very short-term fluctuations.
A characteristic of the model generally?
A characteristic of the model generally?
August 27, 2025 at 12:17 AM
Ferran, wonder what you think of these FNV3 (GDMI) intensity forecasts for Erin. I think it's fascinating how it got the longer-scale (multi-day) changes practically dead on, but couldn't/didn't catch the very short-term fluctuations.
A characteristic of the model generally?
A characteristic of the model generally?
Yesterday I showed Google DeepMind errors for Erin. Thought I'd repeat the analysis for 2025 as a whole, and add the HCCA corrected consensus to the mix. Sample is mainly Erin, of course. GDMI has been impressive so far, especially through 72 h and for intensity. (EMXI not in live NHC FTP data.)
August 26, 2025 at 8:55 PM
Yesterday I showed Google DeepMind errors for Erin. Thought I'd repeat the analysis for 2025 as a whole, and add the HCCA corrected consensus to the mix. Sample is mainly Erin, of course. GDMI has been impressive so far, especially through 72 h and for intensity. (EMXI not in live NHC FTP data.)
For intensity, GDMI again beat everything else through 72 h, beat the consensus at all time periods, but trailed one of the HAFS after 72 h. Again, pretty impressive.
August 26, 2025 at 1:01 AM
For intensity, GDMI again beat everything else through 72 h, beat the consensus at all time periods, but trailed one of the HAFS after 72 h. Again, pretty impressive.
I chose Google Deepmind (GDMI) against a slightly different group of models that I thought were more representative (except no EMXI because it's not in the public decks). For track, GDMI was best through 72 h, beat TVCN at all times, but trailed HAFS after 72 h. Not bad at all.
August 26, 2025 at 12:58 AM
I chose Google Deepmind (GDMI) against a slightly different group of models that I thought were more representative (except no EMXI because it's not in the public decks). For track, GDMI was best through 72 h, beat TVCN at all times, but trailed HAFS after 72 h. Not bad at all.
Last time I did this was 2021-22. The metric is the mean intensity error of the consensus for AL/EP when RI was forecast by any member model or observed at the verifying time. Baseline is the performance for 2015-17. Target was to reduce errors in half, and for those two years we hit the target.
August 17, 2025 at 12:36 AM
Last time I did this was 2021-22. The metric is the mean intensity error of the consensus for AL/EP when RI was forecast by any member model or observed at the verifying time. Baseline is the performance for 2015-17. Target was to reduce errors in half, and for those two years we hit the target.
For NHC's version, if all you get out of it is "storm will likely go here" but also capture the part about "hazardous conditions can occur outside of the cone" and are intrigued enough to get more info then it's done its job, I think. Full description is beneath the graphic on the NHC web site.
August 13, 2025 at 7:43 PM
For NHC's version, if all you get out of it is "storm will likely go here" but also capture the part about "hazardous conditions can occur outside of the cone" and are intrigued enough to get more info then it's done its job, I think. Full description is beneath the graphic on the NHC web site.
Here's what I get for the same 1966-2024 historical period, using 12-h motions from the best tracks. Erin's estimated 20-kt speed earlier today would rank 12th.
August 13, 2025 at 12:54 AM
Here's what I get for the same 1966-2024 historical period, using 12-h motions from the best tracks. Erin's estimated 20-kt speed earlier today would rank 12th.
Not at all surprised that his response to dismal economic news was to fire the head of the BLS. So I guess the next time the cone doesn't match where he wants it to go, he'll just fire the forecasters?
August 1, 2025 at 9:41 PM
Not at all surprised that his response to dismal economic news was to fire the head of the BLS. So I guess the next time the cone doesn't match where he wants it to go, he'll just fire the forecasters?
July 21, 2025 at 1:02 AM
"Aleksander Barkov, let's do it again!"
July 12, 2025 at 4:13 PM
"Aleksander Barkov, let's do it again!"