Validation

India / Telangana / Hyderabad · Benchmark comparison

Is the system actually better?

Validation

Does adaptive blending actually outperform the alternatives? Every source, a simple average and a static ensemble are compared on the same context.

6.75 RMSE · best in class

Event

Region

Variable

Lead time

Forecast error

Root-mean-square error, lower is better.

Simple average9.17
Best single model (BharatFS)7.7
BharatCast adaptive blend6.75

12.4% lower RMSE

than the best individual source (BharatFS) in this context · 16.8% lower than the simple-average baseline.

Benchmark comparison

Seven candidates on the same verification sample.

Lower is better · deterministic demonstration data

RMSE

6.75

adaptive blend

MAE

5.52

adaptive blend

Bias

-0.03

adaptive blend

CSI

0.79

adaptive blend

POD

0.89

adaptive blend

FAR

0.14

adaptive blend

ETS

0.74

adaptive blend

Event replay

Move the slider to replay how the forecast, weights and confidence evolved toward the observation.

T-0
T-48hT-24hT-12hT-6hT-0

Blended forecast

44.4 mm

Observation

44.5 mm

Absolute error

0.1 mm

Confidence

87%

Weights at T-0

BharatFS31%
NCUM27%
GFS14%
GEFS20%
AI Forecast Proxy8%

Context under test

Region

Telangana

Regime

Heavy rainfall

Lead time

24h

Sample

90-day rolling hindcast

Prototype validation uses deterministic demonstration data. Replace with operational hindcast and observation datasets for production deployment. The comparison logic, metrics and baselines are the ones a production system would use.