1What changes
- Activation years: none. Both models meet the rule in 1998, 1999, 2012, 2018, 2019, 2022: 1-in-4.5, as in the framework. Scores against FloodScan are therefore identical (section 5).
- Timing within the season does change. The first day the rule is met moves by 1 day earlier in 1998, 21 days later in 2012, 24 days earlier in 2019, 6 days earlier in 2022 (same day in 1999, 2018). The two large moves are 2012, which the new model reaches 21 days after the old model and after FloodScan's first day over its own 1-in-4 level, and 2019, which it reaches 24 days earlier.
- Forecasts (archive 2016–2022 on both models, leads 0–7 days): the old model's forecasts meet the rule in 2018, 2019, 2022, the new model's in 2016, 2018, 2019, 2022. 2016 is new: a forecast issued 23 Sep put 9 of 10 gauges over their level, while neither model's record nor FloodScan show a 1-in-4 flood that year. The first qualifying forecast moves by 11 days later in 2018, 11 days earlier in 2019, 3 days earlier in 2022.
- Levels. The new model runs about three times higher at these gauges: 1-in-4 levels of 3,462–3,556 m³/s against 1,102–1,117 on the nine main-channel gauges. Its August–November flows follow FloodScan slightly more closely (best correlation 0.73–0.76 against 0.71–0.74), and match it best with no time offset on the main channel, where the old model's best match was offset by three days.
- Same years throughout. Every old-versus-new comparison on this page uses only the years both versions cover: the record 1998–2023 (the calibration years) and, for forecasts, the seasons 2016–2022 that both archives cover completely. What the new model says about 2023–2025 beyond that is in section 6: the rule is not met in 2024 or 2025.
- Gauges. This page keeps the ten endorsed gauges and refits only their levels. A variant page also re-runs the design's gauge selection on the new model and applies the rule to the ten gauges it picks.
hybas_ identifiers. The monitoring reads Google against the endorsed levels (about 1,100–1,120 m³/s on the main channel). The new model's
ordinary wet-season flows sit far above those levels: applied to the new record they would have met the rule in
every one of the 26 seasons 1998–2023 (and in 2024 and 2025 too). The levels have to be replaced by the refitted ones (section 4) when the
API switches, and before that if Google announces a date. Signals that the switch has happened: the warning levels or gauge model
identifiers change for these gauges, or identifiers move from hybas_ to subhybas_ as in the new dataset (not confirmed).2The two versions
| Old model (design) | New model (November 2026) | |
|---|---|---|
| Record (reanalysis) | 1980 – Dec 2023 | Dec 1980 – Dec 2025 |
| Forecast archive (reforecast) | issued Jan 2016 – Jun 2023, leads 0–7 days | issued Jan 2016 – Dec 2025, leads 0–10 days |
| Model | lumped, one model per basin | semi-distributed, routed along the HydroBASINS river network |
| Gauges | the same ten HydroBASINS basins, with identical outlet coordinates | |
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 design's reading of forecast dates applies to both unchanged. Leads 8–10 of the new forecast archive are not used here.
3When the trigger is met
One column per year. FloodScan row: the rarest FloodScan flood level reached in Adamawa that year (3, 4 or 5 for 1-in-3, 1-in-4, 1-in-5; the benchmark the design was scored on). Model rows: the most gauges over their own 1-in-4 level on any one day of August–November; filled where the rule is met (6 or more). Forecast rows: a year in which the forecasts at leads of 0 to 7 days put 6 or more gauges over their level on the same day (the design's reading: the highest forecast value per gauge and day across issues). Years 1998–2023, the years both versions cover; forecasts 2016–2022, the seasons both forecast archives cover completely (hatched otherwise). Hover a cell for details.
| 98 | 99 | 00 | 01 | 02 | 03 | 04 | 05 | 06 | 07 | 08 | 09 | 10 | 11 | 12 | 13 | 14 | 15 | 16 | 17 | 18 | 19 | 20 | 21 | 22 | 23 | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| FloodScan flood year | 3 | 5 | 3 | 5 | 5 | 4 | 3 | 5 | 5 | |||||||||||||||||
| Old model: gauges over | 10 | 10 | 0 | 2 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 10 | 0 | 0 | 0 | 0 | 1 | 9 | 7 | 0 | 0 | 10 | 0 |
| New model: gauges over | 10 | 10 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 10 | 0 | 0 | 0 | 1 | 0 | 9 | 9 | 1 | 0 | 9 | 0 |
| Old model: forecast | ● | ● | ● | |||||||||||||||||||||||
| New model: forecast | ● | ● | ● | ● |
Dates
First day each model's record meets the rule, and first forecast issue date that meets it. The last two columns are the reference the design used for timing: the first day FloodScan crossed its own 1-in-4 level that season, and the FloodScan class of the year.
| Year | Record: old | Record: new | Change | Forecast: old | Forecast: new | Change | FloodScan over 1-in-4 | FloodScan year |
|---|---|---|---|---|---|---|---|---|
| 1998 | 26 Sep | 25 Sep | 1 day earlier | – | – | – | – | 1-in-3 |
| 1999 | 7 Sep | 7 Sep | same day | – | – | – | 18 Oct | 1-in-5 |
| 2012 | 23 Aug | 13 Sep | 21 days later | – | – | – | 23 Aug | 1-in-5 |
| 2016 | – | – | – | – | 23 Sep | – | – | below 1-in-3 |
| 2018 | 13 Sep | 13 Sep | same day | 2 Sep | 13 Sep | 11 days later | 12 Sep | 1-in-4 |
| 2019 | 16 Oct | 22 Sep | 24 days earlier | 2 Oct | 21 Sep | 11 days earlier | – | 1-in-3 |
| 2022 | 3 Sep | 28 Aug | 6 days earlier | 26 Aug | 23 Aug | 3 days earlier | 28 Aug | 1-in-5 |
Against FloodScan's first day over its 1-in-4 level, the first qualifying forecast came 2018: old 10 days before, new 1 day after; 2022: old 2 days before, new 5 days before.
Daily flow at the sixth-highest of the ten gauges, as a share of that gauge's own 1-in-4 level, August to November, on each model's record; the rule is met on days the line is above 100%, and dots mark the first such day. FloodScan: flooded fraction in Adamawa as a share of its own 1-in-4 level. Years shown: every year from 1998 to 2023 in which either model or either forecast met the rule.
4Gauge levels
1-in-4 level at each gauge fitted on the August–November annual maxima 1998–2023 of each model's own record (Weibull plotting position with log-log interpolation, as in the design), and each gauge's best rank correlation with FloodScan over lags of −3 to +14 days (the gauge-selection measure). The old levels are the endorsed ones; the new levels are what the monitoring would need once Flood Hub serves the new model (m³/s).
| Gauge (HydroBASINS) | Outlet | Old level | New level | New / old | Old correlation (lag) | New correlation (lag) |
|---|---|---|---|---|---|---|
| 1120842990 | 9.385, 12.806 | 1,110.8 | 3,461.9 | 3.12x | 0.74 (-3 d) | 0.76 (0 d) |
| 1120843610 | 9.369, 12.852 | 1,101.5 | 3,462.8 | 3.14x | 0.74 (-3 d) | 0.75 (0 d) |
| 1120845060 | 9.331, 12.715 | 1,102.2 | 3,517.3 | 3.19x | 0.73 (-3 d) | 0.76 (0 d) |
| 1120849600 | 9.244, 12.577 | 1,113.7 | 3,518.0 | 3.16x | 0.73 (-3 d) | 0.76 (0 d) |
| 1120848550 | 9.265, 12.494 | 1,106.0 | 3,550.4 | 3.21x | 0.73 (-3 d) | 0.76 (0 d) |
| 1121970280 | 9.315, 12.435 | 1,110.1 | 3,555.8 | 3.20x | 0.72 (-3 d) | 0.76 (0 d) |
| 1120842550 (Wuroboki, quality-verified) | 9.394, 12.369 | 1,114.0 | 3,548.9 | 3.19x | 0.72 (-3 d) | 0.76 (0 d) |
| 1120840700 | 9.444, 12.356 | 1,112.6 | 3,549.1 | 3.19x | 0.72 (-3 d) | 0.76 (0 d) |
| 1120840560 | 9.448, 12.348 | 1,117.1 | 3,549.8 | 3.18x | 0.72 (-3 d) | 0.76 (0 d) |
| 1120840690 (small tributary) | 9.444, 12.340 | 142.7 | 224.6 | 1.57x | 0.71 (+2 d) | 0.73 (+1 d) |
5Performance against FloodScan
As scored in the design: a year counts as a hit when the trigger is met in a FloodScan flood year at the given level, a false alarm when it is met in any other year. Record: 1998–2023. Forecast: the years both forecast archives cover with the season complete.
| FloodScan level | Series | Old model | New model | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| Hits | False | Missed | POD | FAR | Hits | False | Missed | POD | FAR | ||
| 1-in-3 | record, 1998–2023 | 6 | 0 | 3 | 0.67 | 0.00 | 6 | 0 | 3 | 0.67 | 0.00 |
| 1-in-4 | record, 1998–2023 | 4 | 2 | 2 | 0.67 | 0.33 | 4 | 2 | 2 | 0.67 | 0.33 |
| 1-in-5 | record, 1998–2023 | 3 | 3 | 2 | 0.60 | 0.50 | 3 | 3 | 2 | 0.60 | 0.50 |
| 1-in-3 | forecast, 2016–2022 | 3 | 0 | 0 | 1.00 | 0.00 | 3 | 1 | 0 | 1.00 | 0.25 |
| 1-in-4 | forecast, 2016–2022 | 2 | 1 | 0 | 1.00 | 0.33 | 2 | 2 | 0 | 1.00 | 0.50 |
| 1-in-5 | forecast, 2016–2022 | 1 | 2 | 0 | 1.00 | 0.67 | 1 | 3 | 0 | 1.00 | 0.75 |
Other return periods, same 6-of-10 rule
| Level | Old model, years | New model, years |
|---|---|---|
| 1-in-3 | 1998, 1999, 2001, 2010, 2012, 2018, 2019, 2022 | 1998, 1999, 2012, 2016, 2018, 2019, 2020, 2022 |
| 1-in-4 | 1998, 1999, 2012, 2018, 2019, 2022 | 1998, 1999, 2012, 2018, 2019, 2022 |
| 1-in-5 | 1998, 1999, 2012, 2019, 2022 | 1998, 1999, 2012, 2019, 2022 |
| 1-in-6 | 1998, 1999, 2012, 2022 | 1999, 2012, 2022 |
6New model only: 2023–2025
The new record runs two years, and the new forecast archive two and a half seasons, past the old ones. These years have no old-model counterpart and are kept out of every comparison above. Levels are still those fitted on 1998–2023. FloodScan in this benchmark ends in 2023.
| Year | Record: most gauges over on one day | Record: sixth-highest gauge at its peak | Forecasts: most gauges over (leads 0–7) | Rule met |
|---|---|---|---|---|
| 2023 | shared record, see above | 0 of 10 | no | |
| 2024 | 0 of 10 | 90% of its level | 0 of 10 | no |
| 2025 | 1 of 10 | 71% of its level | 1 of 10 | no |
The framework did activate on 7–8 September 2025, under the 2025 version's single-gauge rule on the forecasts of the time. The 2026 rule on the new model is not met in 2025: at most one of the ten gauges went over its level on any day, and the sixth-highest peaked at 71% of its level. The old model's archive ends in 2023, so it cannot be compared for 2025.
7Method
- The design notebook's steps (
exploration/2026/cerf/workflow/06_trigger_definition.ipynb) run on each model's data in turn, inanalysis.pynext to this page: the estimator is copied verbatim; a gauge counts when strictly above its level; forecast valid day = issue day + lead, leads 0–7, best value per day across issues, years whose forecasts reach October. - The old model run here reproduces the endorsed levels to 0.1 m³/s and the endorsed years (record 1998, 1999, 2012, 2018, 2019, 2022; forecast 2018, 2019, 2022).
- Same years for both versions: levels fitted on 1998–2023 for both; activations, timing, FloodScan scores, correlations and the old-levels-on-new-flows count all over 1998–2023; forecasts over the complete seasons 2016–2022.
- The gauge set is the endorsed one; gauge selection is re-run on the new model in the variant page. FloodScan benchmark years are the design's.
- Data: Google Runoff Reanalysis & Reforecast, public Google Cloud bucket
flood-forecasting, old versionmodel_id_8583a5c2_v0, new versiondistributed_model_2026_11(CC-BY-4.0).