Water-level forecasting pipeline
A real-world forecasting pipeline where event timing and hydrological context matter as much as aggregate error.
A hydrological forecasting workflow around TimeGPT, classical baselines, rainfall features, and rolling-origin evaluation.
problem
Hydrological forecasting requires sensitivity to rainfall, upstream stations, event periods, and the operational meaning of early detection.
why it matters
It keeps the portfolio honest about messy, non-math-domain data while still fitting the substrate/evaluation pattern.
maker note
This is applied AI with real-world friction: missing values, stations, flood events, and evaluation choices that matter outside the notebook.
what I built
- • Time-series forecasting workflow
- • Baseline/evaluation comparisons
- • Research framing around sensitivity and event analysis
evidence
- • Published IJFIS TimeGPT water-level paper
- • Follow-up meeting notes around physical loss, station sensitivity, and mid-scale flood analysis
current status
Published foundation-model forecasting paper with ongoing follow-up analysis direction.
next step
Frame the next draft around event detection, physical loss, and sensitivity rather than only raw prediction accuracy.
role
AI forecasting/evaluation contributor.
tags
Public external link pending; this detail page keeps the research narrative stable until the artifact is ready to expose.