← Scenario sandboxFloodLens BD · observed historical data

Real readings. Measured errors.

Loading the archived observations…

Historical research demonstration · not a current warning

This page uses measured river levels from one station in 2024. A Python-trained model predicts the following day's gauge level using only readings available on the input date. It does not validate the six-input scenario sandbox, drought scoring, flood extent, or a nationwide warning service.

Real observations——
24h model MAE—Held-out chronological test
Persistence MAE—Tomorrow equals today's level
Test forecasts—Not independent flood events

Observed river levels and held-out forecasts

● Measured gauge level● Next-day model prediction (test only)

Gaps are left unconnected. Measurements are gauge heights in meters, not river level above normal and not flood depth.

Replay a test forecast



Predicted next-day level—
Observed next-day level—
Absolute error—

This is a historical replay: the target observation is shown for evaluation. It is excluded from that prediction's input features.

Backtest against simple baselines

MethodMAE (m) ↓RMSE (m) ↓

Features: today's gauge level, the one-day change, and the two-day change. Ridge strength is selected on the validation block; the model is then fitted on train + validation. The final test block remains unseen during selection and fitting. Missing days are not interpolated; examples require four consecutive calendar dates, including the target. No confidence interval or calibrated disaster probability is claimed.

Source, reproducibility and limits

The World Bank reproducibility package RR_BGD_2025_385v2 identifies its Bangladesh water-level data as sourced from FFWC/BWDB. The package's analysis script calls the station “Bahurabad.” This page preserves that source spelling; gauge identifiers, datum and official danger thresholds are not present in the two-column data file.

Only one station, one season, and a small test block are available. Errors on this block do not establish performance during unseen flood peaks, other seasons or other locations. No rainfall, soil-moisture observations or drought outcome labels were supplied. The direct FFWC API was inaccessible during this integration; this page does not fetch live readings.

Download observations, predictions and metrics (JSON) · Observed daily levels (CSV)

Public source attribution: FFWC/BWDB via World Bank (2025). Redistribution notice: World Bank BSD-3-Clause license. No endorsement is implied. The source file and analysis script are retained in the source package.