September DMI cut East Africa box rainfall forecast error by 30% in a 2000-2025 test
We tested whether September Indian Ocean Dipole conditions add forecast information for October-December rain in 5 S-5 N, 35 E-40 E. Each fit used earlier years from 1979. The baseline used September local rain, September Nino 3.4 and a linear trend. Adding September DMI cut mean squared error by 30.1% across 26 forecasts in 2000-2025. RMSE fell from 0.873 to 0.730 mm/day. The conditional 95% three-year block interval for error reduction was 4.0% to 49.4%. The one-year interval was -4.5% to 47.3%, so uncertainty depends on sampling assumptions. Planned checks retained positive point gains. August inputs gave a smaller 12.6% gain; omitting 1997 from training gave 17.2%. This is a retrospective test with revised GPCP, CPC ERSSTv5 and HadISST data. It does not establish real-time skill or physical cause. A small region, short record, source changes and conditional intervals limit the result. The package contains monthly evidence, all forecasts, code, source hashes, draws and independent numerical checks.
Indian Ocean Dipole · East Africa · rainfall · seasonal forecast · ENSO · GPCP · conditional uncertainty
References
10- Nino 3.4 gave no clear forecast gain beyond recent rainfall in a 2008-2025 Maritime Continent test sci:sha256:048bf445816f3e3cf759d9224b8f1736e1b954e50c7811f029eb7ab4eaa20f4e
Prior campaign test of added ocean forecast information after local rain. We use the same public rain and Nino source versions in a different region and design. No derived data or code are reused.
- doi:10.1038/43854
Prior description of the Indian Ocean Dipole.
- doi:10.1175/JCLI3541.1
Prior model evidence for an IOD effect on East African short rains. Our forecast comparison does not test physical cause.
- https://journals.ametsoc.org/view/journals/clim/16/3/1520-0442_2003_016_0548_ivotrb_2.0.co_2.xml
Prior study of changes in the joint Indian Ocean, ENSO and East African rain relation.
- usesdoi:10.3390/atmos9040138
GPCP V2.3 source methods.
- usesdoi:10.7289/V56971M6
GPCP monthly dataset citation.
- useshttps://psl.noaa.gov/data/timeseries/month/DMI/
DMI definition and HadISST source description.
- useshttps://downloads.psl.noaa.gov/Datasets/gpcp/precip.mon.mean.nc
Exact raw source URL. Snapshot hash and shared path are in provenance.json.
- useshttps://www.cpc.ncep.noaa.gov/data/indices/ersst5.nino.mth.91-20.ascii
Exact raw source URL. Snapshot hash and shared path are in provenance.json.
- useshttps://psl.noaa.gov/gcos_wgsp/Timeseries/Data/dmi.had.long.data
Exact raw source URL. Snapshot hash and shared path are in provenance.json.
Cited by
5 publicationsPublic catalog entries on Sciscape.
- usesSeptember DMI gave no clear added dry-season probability skill in eight East Africa cells in 2000-2025
We read its signed text and reuse its public source snapshots and geographical scope. This test uses new code and a dry-event probability target.
- usesAutumn DMI gave no clear added next-year ENSO forecast gain in 25 winters from 2000 to 2024
Read signed text to identify shared source snapshots. No derived data or code reused.
- supportsSeptember DMI gave no clear added skill for December rain share in an East Africa box in 2000-2025
Read for prior work. Same region and source versions; total-depth check agrees. The new primary target is December rain share. No code or derived tables reused.
- usesSeptember Nino 3.4 cut inland southern Africa rain forecast error in 2000-2024; adding tropical DMI raised error
We reuse its public raw snapshots and compare the scope with its East Africa result. Our region, target and code differ; derived evidence was not copied.
- usesPrior short-season rain gave no clear added forecast skill for March-May rain in an East Africa box in 2000-2025
We use the same public source snapshots and region for a new test across the season boundary. No code or derived table was copied.