The new Google model and the Gu windows

1What changes

• Activation years. Gu Juba: 2013, 2016, 2018, 2020 on the old model, 2016, 2018, 2020 on the new. Gu Shabelle: 2013, 2016, 2018, 2020 on the old, 2016, 2018, 2020, 2023 on the new. Both windows drop 2013, which the river gauges did not record as a Gu flood season on either river; the Shabelle adds 2023, a severe Gu season on the gauges.
• Timing. Wherever both models meet the rule in the same year, on the record or on the forecasts, the new model meets it later: by 2 to 10 days. The detail per year is in the timing tables below.
• All four windows together (Deyr unchanged on GloFAS version 5, as designed): 8 activations (1-in-3.2) on the old model, 7 (1-in-3.7) on the new; severe years caught 7 and 7 of 7; activations outside a flood season: 2013 and none.
• Levels. The new model runs higher at these points, so the levels rise: Juba points 1.7–2.1 times the old levels, Shabelle points 1.2–1.3 times. Its March–May flows track the SWALIM gauges about as well as the old model on the Juba (median correlation 0.86 → 0.88) and better on the Shabelle (0.79 → 0.89).
• Same years throughout. Every old-versus-new comparison on this page uses only the years both versions cover: the record 1999–2023 (the calibration years) and, for forecasts, the Gu seasons 2016–2023. The new model's 2024 and 2025 are reported separately in section 6.
• Model choice. This page keeps the adopted choice of model per window. A variant page re-runs that choice with the new Google model as a candidate.

2The two versions

Both versions label days the same way (the record for day T equals the forecast issued on day T+1 at lead 0), so the trigger page's reading of forecast dates applies to both unchanged. Leads 8–10 of the new forecast archive are not used here, to keep the action leg as designed.

3Gu Juba: 3 of 4 points over their 1-in-5 level

Each row of the grid is one series, each column one year. River gauges: F a flood season (two gauges over their own 1-in-3 level), S a severe one (two over 1-in-5). Model rows give the most points over their own level on any one day of March to May; a filled cell is a year the rule is met (at least 3). Forecast rows mark a year in which one forecast, at leads of 1 to 7 days, met the rule. Years 1999–2023, the years both versions cover; forecasts 2016–2023, the Gu seasons both forecast archives cover (hatched before). Hover a cell for details.

When the rule is met

First day the model's own record meets the rule, and first forecast issue date that meets it (leads 1–7 days, read as on the trigger page: valid day = issue day + lead). The last two columns are the reference from the river gauges: the day the second gauge crossed its own 1-in-3 level, and how the gauges classed the season.

Daily flow at the point ranked 3rd highest of 4, as a share of that point's own 1-in-5 level, March to May, on each model's record. The rule is met on days the line reaches 100%; dots mark the first such day. Years shown: every year from 1999 to 2023 in which either model or either forecast met the rule.

Levels at each point

1-in-5 level fitted on the March–May seasonal maxima 1999–2023 of each model's own record (Weibull plotting position, as in the trigger analysis) and mean March–May flow (m³/s), and the best-lag rank correlation with the point's SWALIM gauge, computed as in the trigger page's station table.

4Gu Shabelle: 2 of 3 points over their 1-in-6 level

Each row of the grid is one series, each column one year. River gauges: F a flood season (two gauges over their own 1-in-3 level), S a severe one (two over 1-in-5). Model rows give the most points over their own level on any one day of March to May; a filled cell is a year the rule is met (at least 2). Forecast rows mark a year in which one forecast, at leads of 1 to 7 days, met the rule. Years 1999–2023, the years both versions cover; forecasts 2016–2023, the Gu seasons both forecast archives cover (hatched before). Hover a cell for details.

When the rule is met

First day the model's own record meets the rule, and first forecast issue date that meets it (leads 1–7 days, read as on the trigger page: valid day = issue day + lead). The last two columns are the reference from the river gauges: the day the second gauge crossed its own 1-in-3 level, and how the gauges classed the season.

Daily flow at the point ranked 2nd highest of 3, as a share of that point's own 1-in-6 level, March to May, on each model's record. The rule is met on days the line reaches 100%; dots mark the first such day. Years shown: every year from 1999 to 2023 in which either model or either forecast met the rule.

Levels at each point

1-in-6 level fitted on the March–May seasonal maxima 1999–2023 of each model's own record (Weibull plotting position, as in the trigger analysis) and mean March–May flow (m³/s), and the best-lag rank correlation with the point's SWALIM gauge, computed as in the trigger page's station table.

5All four windows together

The allocation is sized on the union of the four windows. Deyr runs on GloFAS version 5 as designed and is the same in both rows (2006, 2023 on the Juba, 2006, 2014, 2019, 2023 on the Shabelle). Severe years are those in which two gauges on a river crossed their own 1-in-5 level in either season (2006, 2014, 2016, 2018, 2019, 2020, 2023).

The windows were calibrated jointly to about 1-in-3 on the old model. On the new model the same settings activate less often (1-in-3.7). Whether to re-tune for the new model is a design decision for the working group; this page only shows the adopted settings carried over.

6New model only: 2024 and 2025

The new record and forecast archive run two years past the old ones, so these years have no old-model counterpart and are kept out of every comparison above. Levels are still those fitted on 1999–2023; the gauge season uses the same gauge levels on the gauges still reporting.

7Method

• Exactly the trigger page's code (scripts/page_additions/somlib.py), run on each model's data in turn: per-point levels by Weibull plotting position with log-log interpolation on March–May maxima 1999–2023; a day counts when at least the required number of points are at or over their own level; forecasts at leads 1–7 days against the same levels.
• Same years for both versions: levels fitted on 1999–2023 for both; activations, timing, scores, correlations and the old-levels-on-new-flows count all over 1999–2023; forecasts over the Gu seasons 2016–2023.
• The river-gauge benchmark is unchanged (two SWALIM gauges over their own level, levels fitted 2000–2023).
• The old model run here reproduces the adopted Gu activation years (2013, 2016, 2018, 2020 on both rivers).
• Data: Google Runoff Reanalysis & Reforecast, public Google Cloud bucket flood-forecasting, old version model_id_8583a5c2_v0, new version distributed_model_2026_11 (CC-BY-4.0).