New Google model: re-running the model choice — Somalia Riverine Flood Trigger

4 of 4

windows where the new Google model tracks the gauges better than GloFAS

1 of 4

windows it carries under the design's caps

3 of 4

windows it carries without the season cap

7 of 7

severe years caught by every option below

1The short answer

Tracking. The new Google model follows the river gauges more closely than GloFAS in all four windows. The change is largest in Deyr, where the old Google model was weak (0.60 and 0.62) and the new one is not (0.87 and 0.87). That weakness is why Deyr went to GloFAS.
The design's search, re-run. With the design's two caps (activate in at most 8 of 25 years; release in at most 8 seasons, counting each Gu or Deyr of a given year in which either river's window activates), the search hands the new Google model only Gu Juba; GloFAS version 5 takes the other three windows. With the year cap alone, the new model also takes Gu Shabelle and Deyr Shabelle. Deyr Juba stays on GloFAS version 5 either way.
Forecast lead, scored only on the days the GloFAS forecast archive holds. Deyr Shabelle: the new Google model met the rule on the same forecast issue as GloFAS in 2019 and 2020, and also caught 2023, which GloFAS missed. Deyr Juba: GloFAS met the rule ahead of the gauges in both flood seasons (2017, 2023); the new Google model missed 2017, was late in 2023 and also met the rule in two seasons without a flood (2019, 2020). In Gu the new Google model reaches the rule later than the old one, but still ahead of GloFAS on the Juba.

Where the search and the design's own principle disagree. Under both caps the search moves Gu Shabelle from Google to GloFAS version 5, which tracks that window worst of the three (0.71 against 0.86 for the new Google model). It does so only because GloFAS activates in 2005, a flood season, which lifts the count from 7 to 8 years. In exchange it gives up Gu Shabelle 2023, a severe season on the gauges; 2023 stays in its record only because the Deyr windows activate in Deyr that year. The design's stated rule was that where the backtest cannot tell models apart, tracking and forecast lead decide. That points to the 7-year option, which is exactly the design on the comparison page: the adopted settings with the new Google model in both Gu windows (1-in-3.7, all severe years, no activation outside a flood season). Which to adopt is for the working group.

2How each model tracks the gauges

The search's own measure, used to break ties: mean best-lag rank correlation of the window's points with the river's reference gauge (Luuq on the Juba, Belet Weyne on the Shabelle), 1999–2023. The comparison page reports a different measure (each point against its own gauge, as in the trigger page's station table), so the values differ; in Gu Juba the two measures rank the old and new Google models the other way round. Bold: best in the window.

Window	Google (old)	Google (new)	GloFAS v4	GloFAS v5
Gu Juba	0.87 (adopted)	0.85	0.82	0.81
Deyr Juba	0.60	0.87	0.84	0.82 (adopted)
Gu Shabelle	0.77 (adopted)	0.86	0.83	0.71
Deyr Shabelle	0.62	0.87	0.67	0.81 (adopted)

3The design search, re-run

For every window the search tries each model, a station return period from 1-in-3 to 1-in-6 and a vote count from two to all-but-one points, with every window activating at least twice. It then takes the combination of the four windows with the most activation years under the caps, and among those the one that catches most severe years, has fewest activations outside a flood season, the fewest unanimous rules, the best tracking and the lowest return periods. Run on the old models it reproduces the adopted design exactly. That check rests on the two caps, not on the ranking: only a handful of combinations pass both caps, and reordering the ranking gives the same pick. The caps are applied together, which is one reading of the design notes; they do not say whether the season cap was applied together with the year cap or afterwards. Under both caps, the new-model pick is the only model assignment with its backtest score.

Option	Gu Juba	Deyr Juba	Gu Shabelle	Deyr Shabelle	Years	Seasons released	Severe caught	Outside a flood season
Adopted design (old Google, GloFAS v4/v5)	Google (old) 3 of 4, 1-in-5	GloFAS v5 3 of 4, 1-in-4	Google (old) 2 of 3, 1-in-6	GloFAS v5 2 of 3, 1-in-5	8 (1-in-3.2)	8	7 of 7	2013
New Google: design caps (8 years, 8 seasons)	Google (new) 3 of 4, 1-in-5	GloFAS v5 3 of 4, 1-in-4	GloFAS v5 2 of 3, 1-in-6	GloFAS v5 2 of 3, 1-in-5	8 (1-in-3.2)	8	7 of 7	none
New Google: years cap only (8 years)	Google (new) 3 of 4, 1-in-5	GloFAS v5 3 of 4, 1-in-4	Google (new) 2 of 3, 1-in-5	Google (new) 2 of 3, 1-in-5	8 (1-in-3.2)	11	7 of 7	none
New Google: 7 years, 8 seasons	Google (new) 3 of 4, 1-in-5	GloFAS v5 3 of 4, 1-in-4	Google (new) 2 of 3, 1-in-6	GloFAS v5 2 of 3, 1-in-5	7 (1-in-3.7)	8	7 of 7	none

Activation years by option (G: released in Gu, D: in Deyr, GD: both). River gauges: F a flood season, S a severe one.

	03	05	06	08	10	13	14	16	17	18	19	20	23
River gauges	F	F	S	F	F		S	S	F	S	S	S	S
Adopted			D			G	D	G		G	D	G	D
Design caps		G	D				D	G		G	D	G	D
Years cap only			D		GD		D	G		G	D	GD	GD
7 years			D				D	G		G	D	G	GD

4Forecast lead

The other half of the design's judgement: which forecast reaches the rule first. For each window, the adopted rule (points and return period as designed) on each forecast archive against that model's own levels, at leads of 1 to 7 days, in the gauge flood seasons of 2016–2023: the first forecast issue that met the rule, and how long before the second gauge crossed its own 1-in-3 level. Every model is scored only on the issue dates the GloFAS forecast archive holds (it is not daily and has no issues from July to September), so each gets the same chances. Its first Deyr issue each year is in early October, so a flood that starts before then (the Shabelle in 2020, gauges over on 1 October) cannot be forecast ahead by any model here. The old Google archive ends in June 2023, so it has no Deyr 2023. The last row of each window lists the years a forecast met the rule in a season the gauges did not record as a flood. The live pipeline reads Google at leads 1–5 (see the comparison page).

Window	Year	Gauges over 1-in-3	Google (old)	Google (new)	GloFAS v4
Gu Juba 3 of 4, 1-in-5	2016	1 May	1 May (same day)	8 May (7 d after)	11 May (10 d after)
	2018	17 Apr	15 Apr (2 d before)	18 Apr (1 d after)	4 May (17 d after)
	2020	22 Apr	24 Apr (2 d after)	29 Apr (7 d after)	1 May (9 d after)
	2023	25 Mar	–	–	–
	met the rule outside a flood season		none	none	none
Deyr Juba 3 of 4, 1-in-4	2017	7 Nov	–	–	19 Oct (19 d before)
	2023	25 Oct	no archive	28 Oct (3 d after)	21 Oct (4 d before)
	met the rule outside a flood season		2019	2019, 2020	none
Gu Shabelle 2 of 3, 1-in-6	2016	11 May	1 May (10 d before)	9 May (2 d before)	8 May (3 d before)
	2018	7 May	27 Apr (10 d before)	1 May (6 d before)	11 May (4 d after)
	2020	5 May	22 Apr (13 d before)	2 May (3 d before)	–
	2023	17 Apr	–	6 May (19 d after)	–
	met the rule outside a flood season		none	none	none
Deyr Shabelle 2 of 3, 1-in-5	2019	14 Oct	2 Oct (12 d before)	2 Oct (12 d before)	2 Oct (12 d before)
	2020	1 Oct	–	2 Oct (1 d after)	2 Oct (1 d after)
	2023	6 Nov	no archive	4 Nov (2 d before)	–
	met the rule outside a flood season		none	none	none

5Method

Candidates exactly as in scripts/model_selection.py (models, return periods 1-in-3 to 1-in-6, two to all-but-one votes, tracking against the reference gauge, every window at least twice), judged on the two-gauge benchmark of scripts/envelope_search.py. GEOGloWS is left out, as in the adopted design.
The two caps reconstruct the design's two decisions: activate no more often than 1-in-3 (8 of 25 years), and release in no more than 8 seasons, each Gu or Deyr of a given year counted once whichever river activates (the change to Deyr Shabelle on 18 September 2026, which took the adopted design from 10 such seasons to 8). With both, the search returns the adopted design on the old models, which the script asserts. The search is written out in scripts/google_model_selection.py; the pick of model_selection.choose is not used because it has no tracking tie-break and depends on candidate order.
Dollow is the Juba main-stem gauge, as live; the design ran on the Dawa-branch point and the reconstruction still returns the adopted design.
Any window on GloFAS version 5 (Deyr in the adopted design, and Gu Shabelle too in the capped pick) is fitted on version 5 while the live forecast is version 4: see the GloFAS version switch.

Generated 4 November 2026 by scripts/build_google_model_selection_page.py (numbers in scripts/google_model_selection.py) in ds-aa-som-floods. Back to the comparison page.