Why the UK forecast flipped: a storyboard, working backwards

How to read a Hovmöller for packets

• Individual troughs and ridges move east slowly; here the phase speed is about 7–8 m/s.
• Energy moves along the packet at the group speed, here about 25 m/s, building new troughs and ridges downstream.
• In a Z Hovmöller that looks like a stepped, diagonal band of strong anomalies. The v250 envelope shows it as one band.

The packet started near 170°E on 13 Oct, peaked at 45 m/s near 152°W on 15 Oct and reached the UK on 19 Oct.

Trough along 75–90°W

Does the UK outcome depend on this? (14 Oct run, members split)

• Z500 at 45°N, 64°W on 18 Oct: r = −0.73. A trough parked there means a UK high.
• Packet strength over 90–50°W on 18 Oct: r = +0.65. A stronger packet means a UK high.
• A shuffle test puts the 5% significance threshold at about |r| = 0.28 for a single pre-chosen box, and 0.56 for the strongest point anywhere in the hemisphere. Both values clear both.

Box means (mm/day): best estimate · 13 Oct run · 16 Oct run

250 hPa outflow on 18 Oct 00z was about 30% weaker in the 13 Oct run: 6.1 against 9.4 (SE US) and 8.9 against 12.1 (east coast), in 10⁻⁶ s⁻¹.

• A fingerprint of the trough (14 Oct members): a coastal rain band on 18–19 Oct goes with the flip (r = −0.52); rain concentrated over the Maritimes on 17–18 Oct goes with the block (r = +0.55).
• Timing points to the trough first: the 13 Oct run's height errors over North America were already growing on 15–16 Oct, while its rainfall was still about right on 16–17 Oct.

Latent-heating test, 14 Oct run (50 members)

"Ahead" = the downstream half of a disc around each member's wave; results hold for radii of 1,000, 1,500 and 2,000 km and for both 12 h windows tested. Shuffle-test 5% threshold about |r| = 0.28.

The real track

• 13 Oct: a weak feature in the NE Pacific (44°N, 153°W) strengthens through the Gulf of Alaska.
• 14 Oct: comes ashore on the BC coast (53°N, 133°W at 00z) and crosses the Rockies into north-east BC and Alberta.
• 15 Oct: turns south-east over Saskatchewan (49°N, 104°W at 12z).
• 16 Oct: digs through the Dakotas, Iowa and Missouri.
• 17–18 Oct: reaches Tennessee and North Carolina (36–38°N), strengthening from about 14 to 28 ×10⁻⁵ s⁻¹. This is the trough of frame 5.

Before 13 Oct the trace leads back, tentatively, to an upper low north-east of Hawaii (26–29°N, 165–170°W) on 10–11 Oct. The feature nearly vanished on 12 Oct, so that link is uncertain.

Step 1 · 13 Oct run: right for two days, then the wrong track

Within 110–230 km up to 15 Oct 00z. After the Rockies crossing the forecast wave stays 3–6° too far north and lifts out north-east. The error sits on the wave and the ridge immediately behind it (frame 8).

Step 2 · 14 Oct run: right track, not deep enough

• Its members follow the real path closely (16 Oct 00z: 45.8°N against a real 45°N).
• At the trough base on 17 Oct 12z they average 22.9 against a real 28.4 ×10⁻⁵ s⁻¹; only one member reaches 29.1.

r = −0.50 between strength and the UK week (5% threshold about 0.27). The latent-heating test in frame 6 finds no sign that rain drove this deepening, so the missing strength looks dynamical: the upper-level jet and wave structure, which 1° open-data fields cannot resolve well.

Why the flip was gradual

• 13 Oct and earlier: wrong track → no dig → Greenland block → dry UK.
• 14 Oct: right track, too weak → partial flip (UK +10 → +5 hPa).
• 16–17 Oct: the dug, deepened wave is in the analysis itself → full flip.

The mechanism in the 13 Oct members

• Stronger ridge behind the wave → stronger northerly steering: r = −0.43 to −0.68.
• Stronger northerly steering → further south 24 h later, allowing for starting latitude: partial r +0.39 to +0.68; +0.44 to +0.81 across three ring sizes.
• The run's ridge was 8–20 m weaker than analysed and its steering 0.5–3 m/s less northerly. The 14 Oct run had both about right, and its track was right.
• Balloons on 15 Oct 12z: 250 hPa northerlies of −52 m/s at Stony Plain and −47 m/s at Great Falls, against −44 and −38 m/s in the 13 Oct forecast.

The missing northerlies were close to the wave

What the low was

• A cut-off low from the north, not the subtropics: a short wave over the Yukon and SE Alaska on 9 Oct that dropped down the BC and Oregon coast (10–11 Oct), cut off over the Great Basin (12–13 Oct), crossed Arizona and the Four Corners on 14 Oct and reached Colorado and Nebraska on 15 Oct.
• Compact and strong: vorticity about 25 ×10⁻⁵ s⁻¹, closed 500 hPa low near 5,620 m on 14 Oct.
• It phased with the northern wave: the gap closed from 2,400 km (14 Oct) to 893 km (15 Oct 12z); one merged feature by 16 Oct 06z, which then deepened from about 16 to 28 ×10⁻⁵ s⁻¹ as it dug into the south-east US. Phasing combines the two troughs' vorticity into one larger, deeper trough, which fits the trough extending further south.

Sensitivity tests across members

13 Oct 00z run: the short wave

Strength is relative to each method's own analysis value (1.00 = as analysed). Tracking starts at 13 Oct 12z for both systems, because at 13 Oct 00z the wave was too weak in height-derived vorticity to track.

• 14 Oct 00z run: both systems follow the real path (AIFS 111 km off on 15 Oct 12z, IFS 264 km), and both under-deepen the wave by 17 Oct 12z (0.79 and 0.80 of the analysed strength).

Trough, packet and verification

Tropical cyclones near North America Ruled out

• Atlantic / Caribbean: the only storm was Isaias (09L), at 976 hPa on the northern Gulf coast on 10 Oct. It was inland and decaying by 11 Oct. No run from 11 to 17 Oct tracks any Atlantic system.
• Eastern Pacific: Simon weakened off western Mexico; Trudy deepened to about 977 hPa off southern Mexico on 16–18 Oct and moved north-west along the coast. There is no sign either was involved.

Tropical convection and MJO Ruled out as the trigger

• Forecast tropical OLR for 19–26 Oct barely changed between runs (3–6 W/m² by sector, against an El Niño-like background of ±30).
• An MJO-like eastward signal is visible, with a new phase over the western Indian Ocean around 13–19 Oct. This is a visual read only; the RMM index was not computed.

Caveats

• The analysis uses the medium-range ensemble (0.25°, thinned to 1°), not the extended-range product. They agree on the flip.
• "Best estimate" rainfall is the HRES day-1 forecast, not observations. Jet speeds are geostrophic.
• The data show where analyses and forecasts diverged, but not which observations fed each analysis.
• Correlations across members show association, not cause. Rain and 700 hPa ascent stand in for latent heating.
• AIFS short-wave tracking uses vorticity derived from 500 hPa heights (checked against the true-vorticity track on the analyses).
• Observations are radiosondes only (IGRA2, quality flags not applied); aircraft and satellite data are not in the open record. Steering uses the 500 hPa geostrophic wind at a single level.

Changed since the first version

• Koguma demoted from "possible root cause" to context: the packet it seeded was forecast well.
• The chance level for hemisphere-wide correlations is about 0.46–0.56, not 0.4. Only the strongest links are kept.
• Koguma's pressure is quoted as the ensemble mean (about 972 hPa), not the deepest member (953 hPa).
• Short wave traced (frame 7): the "northern stream" origin is now a tracked feature with a two-step forecast error.
• Latent-heating check (frame 6): heating did not drive the deepening. This replaces the earlier "very likely reinforced the split".
• AIFS-ENS cross-check (frame 10): the machine-learnt ensemble made the same error.
• Observations and steering (frame 8): no large missed feature at landfall; the error grew over land around the wave.
• Four Corners low (frame 9): real, well forecast, and a phasing partner for the northern wave; not the source of the error.

Possible next steps

• AIFS initial-condition experiment (priority). Re-run ECMWF's open-weights AIFS model from the 13 Oct 00z analysis in several versions: unchanged; with the Four Corners low weakened or removed; with the northern short wave deepened; and with the ridge behind it sharpened. If the "sharper ridge" or "deeper wave" versions flip the UK week and the "no low" version blocks it, that would turn the associations in frames 7–9 into tested cause and effect. Practical on a GPU workstation; several hours per forecast on a laptop CPU.
• Deep-layer steering and potential-vorticity diagnostics around the short wave on 14–16 Oct with full-resolution data, to separate steering from the wave's own development.
• Examine the jet streak and upper-level wave structure around the short wave on 16–17 Oct with full-resolution data, to find what the 14 Oct run missed.
• Compare the ECMWF MJO (RMM) forecast between the 13 and 16 Oct runs.