GloFAS version switch — Somalia Riverine Flood Trigger

v4.5

operational GloFAS, EWDS, 14 Sep 2026

2.3x

median v4 / v5 ratio of the Deyr action levels

8 → 12

envelope activations 1999–2023, design → live

7 → 6 of 7

severe years caught, design → live

What is live

Checked on 14 September 2026. The EWDS forecast dataset offers one operational system version plus legacy versions 2.1 and 3.1; there is no legacy version 4 stream, which there would be if version 4 had been retired. The EWDS reanalysis dataset labels version 4.0 Operational and version 5.0 Pre-operational. ECMWF's latest-release page names v4.5 (16 April 2026), and CEMS has announced that v5.0 will become operational after pre-operational testing with v4 run in parallel for a period. The operational GRIB carries no version string, only process identifiers (generatingProcessIdentifier 5, backgroundProcess 21), which the pipeline records with every row.

The monitoring pipeline therefore uses the GloFAS v4 levels below. It checks the EWDS catalogue every day and fails, rather than emailing, the day a version 4 entry appears under Legacy Versions or the GRIB process identifiers change. Switching is one constant (GLOFAS_OPERATIONAL in src/monitoring/config.py); both level sets are frozen in src/monitoring/thresholds.json.

The two climatologies

Version 4 runs 2.3 times higher than version 5 on the Deyr seasonal maxima at these points, and nearly flat along each river where version 5 decreases downstream. The versions share the river network here, so every comparison is at the same cell. A level fitted on one version is meaningless on the other.

v4 vs v5 seasonal maxima — Deyr seasonal maximum at each monitored point, 1999–2023, version 4 against version 5. Dashed: equal; thin grey: version 4 twice version 5.

Deyr 2023 on both versions — Deyr 2023, the largest flood in the record, at the two reference gauges on each version's reanalysis with each version's own Deyr level (1-in-4 at Luuq, 1-in-5 at Belet Weyne). At Luuq both versions cross their own level; at Belet Weyne version 5 crosses and version 4 stays below its level through the largest flood on record, which is why no version 4 rule catches Deyr 2023 on the Shabelle.

Deyr action levels, both versions

Levels at each point's Deyr return period, fitted on the Deyr seasonal maxima 1999–2023 of each version's own reanalysis (Weibull plotting position, as in the trigger analysis). The live column is highlighted. The last column is the readiness level: the same version 4 reanalysis level at the window's return period, capped at 1-in-5 (READINESS_RP_CAP in src/monitoring/config.py), so in Deyr it equals the live action level.

River	Point	RP	v5 level (design)	v4 level (live)	v4 / v5	v5 Deyr mean	v4 Deyr mean	Readiness level (v4 reanalysis, leads 8–12 d)
Juba	Dollow	1-in-4	601	1,736	2.89x	259	890	1-in-4: 1,736
Juba	Luuq	1-in-4	973	1,734	1.78x	320	895	1-in-4: 1,734
Juba	Bardheere	1-in-4	966	1,730	1.79x	328	906	1-in-4: 1,730
Juba	Bualle	1-in-4	871	1,725	1.98x	313	914	1-in-4: 1,725
Shabelle	Belet Weyne	1-in-5	477	1,109	2.33x	149	471	1-in-5: 1,109
Shabelle	Bulo Burti	1-in-5	389	1,109	2.85x	116	476	1-in-5: 1,109
Shabelle	Jowhar	1-in-5	351	1,159	3.30x	104	487	1-in-5: 1,159

Gu readiness levels (GloFAS carries readiness only in Gu; Google carries the action leg), version 4 reanalysis at 1-in-5: Dollow 2,177, Luuq 2,176, Bardheere 2,175, Bualle 2,240, Belet Weyne 663, Bulo Burti 645, Jowhar 647 m³/s. On the version 5 reanalysis the same points sit at 783, 867, 822, 748, 462, 352, 318.

What the Deyr windows do on each version

Each window's adopted rule replayed on each version's own reanalysis with levels fitted on that version (the two consistent columns), and with the levels crossed over (the two inconsistent columns, which is what applying the design levels to the live forecast would have meant). Judged against the two-gauge benchmark: a flood season has two of the river's gauges over their own 1-in-3 level, a severe one two over 1-in-5, levels fitted 2000–2023.

Window	v5 levels on v5 flows design	v4 levels on v4 flows live	v5 levels on v4 flows the mismatch	v4 levels on v5 flows
Deyr Juba rule 3 of 4 over 1-in-4	2 (1-in-13.0) 2006, 2023 severe 2/3, outside flood years 0	6 (1-in-4.3) 2001, 2005, 2014, 2015, 2017, 2023 severe 2/3, outside flood years 3	25 (1-in-1.0) 1999, 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010, 2011, 2012, 2013, 2014, 2015, 2016, 2017, 2018, 2019, 2020, 2021, 2022, 2023 severe 3/3, outside flood years 20	1 (1-in-26.0) 2023 severe 1/3, outside flood years 0
Deyr Shabelle rule 2 of 3 over 1-in-5	4 (1-in-6.5) 2006, 2014, 2019, 2023 severe 4/5, outside flood years 0	5 (1-in-5.2) 2010, 2013, 2014, 2019, 2020 severe 3/5, outside flood years 2	25 (1-in-1.0) 1999, 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010, 2011, 2012, 2013, 2014, 2015, 2016, 2017, 2018, 2019, 2020, 2021, 2022, 2023 severe 5/5, outside flood years 19	0 (never) none severe 0/5, outside flood years 0

Applying the design (v5) levels to the live v4 flows would have activated Deyr Juba in 25 and Deyr Shabelle in 25 of 25 years: the levels sit far below version 4's ordinary Deyr flows. This is the failure the version guard exists to prevent.

The envelope

The Gu windows run on Google Flood Hub and are unchanged. The union of all four windows, which is what the allocation is sized on, on the design and on the live Deyr levels:

Deyr levels	Activations	Return period	Years	Severe caught	Severe missed	Outside flood years
Live: v4 levels on v4 flows in Deyr	12	1-in-2.2	2001, 2005, 2010, 2013, 2014, 2015, 2016, 2017, 2018, 2019, 2020, 2023	6 of 7	2006	2001, 2013, 2015
Design: v5 levels on v5 flows in Deyr	8	1-in-3.2	2006, 2013, 2014, 2016, 2018, 2019, 2020, 2023	7 of 7	none	2013

activation strip — Activation years by window and version. Shaded columns are gauge flood seasons on either river (darker: severe). Gu rows are identical in both views.

Does another level restore the design on version 4?

For the two Deyr windows on version 4, the activation record at every return period the record supports and at the adopted vote count and its neighbours. The adopted rule shape is highlighted. Read this before deciding whether the Deyr rules need re-tuning for the live version: the analysis page's calibration searched the envelope jointly, and any change here should go back through that search.

Window	Level	Votes	Activations	RP	Severe	Outside flood years	Years
Deyr Juba	1-in-3	2 of 4	8	1-in-3.2	3/3	4	2001, 2005, 2006, 2011, 2014, 2015, 2017, 2023
Deyr Juba	1-in-3	3 of 4	8	1-in-3.2	3/3	4	2001, 2005, 2006, 2011, 2014, 2015, 2017, 2023
Deyr Juba	1-in-3	4 of 4	7	1-in-3.7	2/3	4	2001, 2005, 2011, 2014, 2015, 2017, 2023
Deyr Juba	1-in-4	2 of 4	6	1-in-4.3	2/3	3	2001, 2005, 2014, 2015, 2017, 2023
Deyr Juba	1-in-4	3 of 4	6	1-in-4.3	2/3	3	2001, 2005, 2014, 2015, 2017, 2023
Deyr Juba	1-in-4	4 of 4	6	1-in-4.3	2/3	3	2001, 2005, 2014, 2015, 2017, 2023
Deyr Juba	1-in-5	2 of 4	5	1-in-5.2	2/3	2	2005, 2014, 2015, 2017, 2023
Deyr Juba	1-in-5	3 of 4	5	1-in-5.2	2/3	2	2005, 2014, 2015, 2017, 2023
Deyr Juba	1-in-5	4 of 4	2	1-in-13.0	1/3	0	2017, 2023
Deyr Juba	1-in-6	2 of 4	4	1-in-6.5	2/3	1	2014, 2015, 2017, 2023
Deyr Juba	1-in-6	3 of 4	3	1-in-8.7	1/3	1	2015, 2017, 2023
Deyr Juba	1-in-6	4 of 4	2	1-in-13.0	1/3	0	2017, 2023
Deyr Shabelle	1-in-3	2 of 3	8	1-in-3.2	4/5	4	2006, 2007, 2010, 2013, 2014, 2017, 2019, 2020
Deyr Shabelle	1-in-3	3 of 3	6	1-in-4.3	3/5	3	2007, 2010, 2013, 2014, 2019, 2020
Deyr Shabelle	1-in-4	2 of 3	6	1-in-4.3	3/5	3	2007, 2010, 2013, 2014, 2019, 2020
Deyr Shabelle	1-in-4	3 of 3	4	1-in-6.5	2/5	2	2010, 2013, 2019, 2020
Deyr Shabelle	1-in-5	2 of 3	5	1-in-5.2	3/5	2	2010, 2013, 2014, 2019, 2020
Deyr Shabelle	1-in-5	3 of 3	3	1-in-8.7	2/5	1	2013, 2019, 2020
Deyr Shabelle	1-in-6	2 of 3	4	1-in-6.5	2/5	2	2010, 2013, 2019, 2020
Deyr Shabelle	1-in-6	3 of 3	2	1-in-13.0	1/5	1	2013, 2019

No version 4 rule reproduces the design. On the version 4 record the adopted Deyr rules activate 6 and 5 times (design: 2 and 4) and the envelope moves from 1-in-3.2 to 1-in-2.2. Severe Deyr seasons that version 4 never registers at any level or vote count tested: Juba none; Shabelle 2023. Raising the level trims the activations outside flood years but does not recover those seasons, so the Deyr windows cannot simply be re-levelled on version 4: the working group has to decide between accepting weaker Deyr coverage on the live version, or waiting for version 5 to go operational and running Deyr on it as designed.

Benchmark by window (two gauges, levels fitted 2000–2023):

Window	Flood years (1-in-3)	Severe years (1-in-5)
Gu Juba	2005, 2010, 2016, 2018, 2020, 2023	2018, 2020, 2023
Deyr Juba	2006, 2008, 2014, 2017, 2023	2006, 2014, 2023
Gu Shabelle	2003, 2005, 2010, 2016, 2018, 2020, 2023	2016, 2020, 2023
Deyr Shabelle	2006, 2008, 2014, 2019, 2020, 2023	2006, 2014, 2019, 2020, 2023

Also live, also different from the design

Dollow's Google point is not the design's. The design used HYBAS gauge hybas_1121038740, which the live Flood Hub API does not serve (404) and which turns out to be the Dawa branch (31 m³/s mean). Since 15 September 2026 Dollow reads hybas_1121039440, the Juba main stem 8 km east-south-east (142 m³/s mean), which the API serves. Against the SWALIM Dollow gauge it tracks as well as the design point did in Gu (Spearman 0.86), and the Gu Juba rule gives the same activation years with either point. The Gu levels for Dollow on this page and in the pipeline are fitted on the main-stem gauge; the trigger pages still show the design point's Dollow numbers until rebuilt.
Google's live horizon. The API returns daily values from two days before the issue to five days after it, so the action leg reads Google at leads 1–5, not 1–7 as in the archive.
Readiness reads the same version 4 reanalysis levels as the action leg, at the window's return period capped at 1-in-5, so the version switch applies to both legs alike.
When version 5 goes live, flip the constant, regenerate this page and the design comparison holds in reverse: the v5 columns above become the live ones. The v5 reanalysis extends to mid-2026 on EWDS, so the levels can also be refitted on a longer record at that point.