C
CI-EWS·CARBETA
Log out
THE ENGINE

CI-EWS Malaria Model

A LightGBM (Poisson, h4) forecaster of confirmed malaria cases at district × month, trained on real CAR surveillance + climate indicators, with conformal prediction intervals — chosen by a head-to-head bake-off and evaluated on a 10-month hold-out.

00Deployed

Live forecast — LightGBM (Poisson, h4)

Confirmed OPD malaria cases per district × month. Forecast for 2025-12 (data through 2025-12); ciews-malaria v0.3.0.

WIS
—
MAE
—
median abs. error
MASE
—
95% coverage
—
interval calibration
National forecast · 2025-12
238,158
cases · 95% 0–1,095,575
National reported cases — trailing 24 months
— reported cases
Per-district forecast · 2025-12
DistrictForecast95% interval
Bangassou13,8950–38,393
Bangui I13,3210–37,819
Bambari13,1300–37,628
Bangui Ii11,4950–35,993
Carnot-Gadzi11,3950–35,892
Alindao-Mingala11,3020–35,799
Paoua10,5870–35,085
Bossangoa10,0360–34,534
Kemo9,5390–34,037
Bangui Iii9,5020–33,999
Bimbo8,8480–33,346
Bouar -Baoro7,9400–32,438
Sangha-Mbaere7,3510–31,849
Kembe-Satema6,5460–31,044
Berberati6,1330–30,631
Mbaiki5,8840–30,382
Baboua-Abba5,7940–30,292
Bossembele5,7470–30,245
Bozoum-Bossemptele5,4820–29,980
Ouango-Gambo5,4130–29,910
Nana-Gribizi5,3910–29,888
Haute-Kotto5,3150–29,813
Begoua5,1880–29,685
Boda5,0310–29,528
Mobaye-Zangba4,9220–29,419
Kouango-Grimari4,6010–29,099
Nanga-Boguila4,3070–28,804
Bamingui-Bangoran4,2970–28,794
Bocaranga-Koui3,8560–28,354
Batangafo-Kabo3,5180–28,015
Ngaoundaye3,1630–27,661
Haut-Mbomou2,5280–27,026
Vakaga2,3760–26,873
Gamboula2,3590–26,857
Bouca1,9660–26,463
01Registry

Every model that's deployed

Straight from CI-EWS Studio's MLflow registry — 5 registered models, one per disease, each with the exact version live right now, when it was trained, and how it scored.

ModelDeployedMASEBeats naiveTrainedStatus
Malaria model — deployed detail
v1
ciews-malaria
lightgbm
21 Jul 2026 · 09:49
ml-engineer
month
1330 train / 340 test rows
14
VersionTrainedMAEWISCoverageAlert F1
v1active21 Jul 2026 · 09:4920731672100%—
02Comparative study

How LightGBM (Poisson, h4) won the bake-off

Every model saw identical features (climate indicators + autoregressive case history) and the same 10-month hold-out. WIS (Weighted Interval Score, lower is better) is the primary metric; skill % is improvement over the seasonal-naive baseline.

ModelWISMAEMASE95% cov.Skill vs naive
LightGBM (Poisson, h4)DEPLOYED
1671.52073.00.34589%+45%
Seasonal Naive (baseline)
1963.63796.50.63299%—
WIS lower is better; skill % is the WIS improvement over seasonal-naive. All families were trained by CI-EWS Studio on identical data and hold-out — this table is regenerated from the Studio/MLflow runs, not hand-authored.
03Data profiling

What the model was trained on

Real CAR surveillance + climate indicators at district × month grain. Window 2022-01 → 2025-12.

Observations
1,670
35 districts × 48 months
Geography
35
districts nationwide
Time span
48 mo
2022-01 → 2025-12
Features
14
selected from 23 candidates
Target — confirmed malaria cases / district-month
5,886
median
7,392
mean
393,497
max
0%
zero months
Feature groups (14 total)
autoregressive (case lags)
7
climate
4
calendar / seasonal
3
04Design

Demand forecasting model

A planned model, not a shipped one — this documents what exists today, what's a placeholder, and what's still needed.

ILLUSTRATIVE — DESIGN SPEC
This model does not exist yet. Two of the six pieces below are real and already shipped (the case forecast and current stock feeds); the conversion step and output shape are running as an explicitly-tagged placeholder preview on the Supply Chain page today; the lead-time, safety-stock and uncertainty-propagation logic are not built at all.
Input
Case forecast (real, exists today)
The deployed LightGBM model already produces a per-district, next-month point forecast with a 95% conformal interval — this is the exact input the demand model will consume.
LIVE
Input
Current stock (real, exists today)
RDT on-hand and ACT/AL courses dispensed per district, read live from HMIS — the exact input for a gap calculation.
LIVE
Conversion
Cases → commodity demand (placeholder today)
Supply Chain's preview uses one flat national ratio (RDTs/ACT-AL per case). The real model needs this calibrated per district and per commodity, since consumption per case varies with testing algorithm, referral patterns, and stockage practice.
PLACEHOLDER
Logic
Lead time & safety stock (not built)
Procurement and distribution take weeks. The real model needs a reorder-point policy: order when projected stock will fall below a safety buffer before the next resupply cycle completes — not just a same-month gap.
NOT BUILT
Logic
Uncertainty propagation (not built)
Today's preview uses only the case forecast's point estimate. The real model should propagate the forecast's 95% interval through to a demand range, so a procurement decision reflects how confident the case forecast actually is.
NOT BUILT
Output
District readiness + national summary (preview shipped)
National and per-district readiness gap, refreshed monthly alongside the case forecast — the output shape is already live on Supply Chain, just running on placeholder logic until the pieces above are built.
PLACEHOLDER
05Benchmarks

Where it stands

Real hold-out comparisons plus the published CAR baselines worth comparing against. Green = cleared. Amber = open comparison.

Open
Seasonal-naive
Malaria
Same month last year — the floor every model must beat.
Open
MASE < 1
Malaria
Mean absolute scaled error below 1 beats seasonal-naive.
Open
95% PI coverage
Malaria
Prediction intervals should cover ~95% of actuals.
Open
Rubuga 2024 (DLNM)
Malaria
Published DLNM across all 35 districts — the CAR malaria baseline to compare against at district grain.
Open
Semakula 2020 (BYM)
Malaria
Bayesian spatial model — the reference for district-level evaluation.
Where every number comes fromSource Catalog -- Every feed, its freshness, completeness, and data lineage.→