Tuesday, January 27, 2015

Don't Blame The Models

I happened to notice this news story (only excerpts below) appearing at a variety of online sites this morning (1/27/15).
MOUNT HOLLY, N.J. (AP) — A National Weather Service official says the agency will evaluate its storm modeling after a storm that was predicted to dump a foot or more of snow on many parts of New Jersey and the Philadelphia region delivered far less than that. ... ... says the storm tracked a bit to the east of what forecasting models predicted. ... ... says the agency will evaluate what happened to see how it can do better in the future.
I am a big opponent of the phrase confidence is high because models have converged on a solution. Red alarm bells should go off any time a forecaster starts to use that phrase. Confidence should be high because alternatives have been ruled out, not because a certain number of models have converged on a small subset of the universe of still plausible scenarios.
But yesterday's storm was not even a case of all models having converged on a solution. I remembered seeing the word outlier in yesterday's forecast discussions. This morning I searched for that word in the Mount Holly discussions, but discovered that outlier had been last used by that office on Sunday afternoon. Instead it was in two of yesterday's discussions for New York, NY, where that word appeared.
New York, NY 10:05 AM [Monday] NWP GUIDANCE HAS SHOWN A LITTLE MORE SPREAD THAN ONE WOULD LIKE TO SEE AT SUCH A SHORT RANGE. MUCH OF THE 00Z GUIDANCE LOWERED QPF AMOUNTS FOR MUCH OF THE REGION...WHICH WOULD MEAN LOWER SNOW TOTALS. HOWEVER... ... DID NOT WANT TO MAKE DRASTIC CHANGES TO THE EXPECTED SNOW AMOUNTS AND HEADLINES WITH JUST ONE MODEL CYCLE. IN FACT... ... THE 00Z GFS APPEARS TO BE A FAST NE OUTLIER.
(The omitted sections in the above quote, after however and in fact, are where the forecaster's confidence was being elevated by model output.)
New York, NY 1:34 PM [Monday] THE STORM APPEARS ON TRACK. THE 12Z GFS IS AN OUTLIER FROM THE 12Z NAM AND 12Z ECMWF. HAVE UPPED SNOWFALL TOTALS ACCORDINGLY WITH MOST AREAS IN THE 20-30 RANGE FOR THIS EVENT.
Nobody wants to hear a forecaster be wishy-washy. People want to hear your best shot. At the same time it is possible to communicate what could go wrong. But instead of acknowledging plausible alternatives, the messages to the public (e.g., upgrading watches to warnings 18 hours before the start of the event) were indicating that forecaster confidence was exceeding record highs.
Don't blame the models. Blame policies, procedures, sloppy communications and training.