AI-driven cash-flow prediction & risk-flagging for rural micro enterprises. NABARD Hackathon @ Global FinTech Fest 2026 — Round 2 prototype.
Forecasts monthly net cash flow 1–6 months ahead per enterprise as P10/P50/P90 quantile bands, converts the band into P(deficit) per future month, and turns that into a tiered early-warning layer with plain-language drivers and field actions. Three cohorts: kharif farmer, dairy SHG member, kirana trader.
Holdout results (6-month, months 37–42): MAE ₹3,393 vs ₹5,987 seasonal baseline (43% better) · flags 88% precision / 79% recall · 76% P10–P90 coverage.
python ml/pipeline.py # simulate panel → train quantile GBMs → export JSON (optional: exports are committed)
cd frontend && npm install
npm run dev # open http://localhost:5173The pipeline exports static JSON artifacts to frontend/public/data/, so the
dashboard needs no backend or retraining at request time — exactly how a
production deployment would serve nightly-batch scores to field devices.
flowchart LR
A[Signal layer<br/>UPI proxies · SHG deposits<br/>mandi price · rainfall dev] --> B[State layer<br/>lags · rolling stats · fatigue<br/>seasonality encoding]
B --> C[Forecast layer<br/>per-horizon quantile GBMs h=1..6<br/>P10 / P50 / P90 bands]
C --> D[Decision layer<br/>P deficit per month · tiered flags<br/>plain-language drivers]
D --> E[Delivery layer<br/>lender dashboard · watch-list<br/>liquidity calendar]
| View | Purpose |
|---|---|
| Overview | Portfolio KPIs, state stress index, cohort mix, top-risk enterprises |
| Watch-list | Tiered flag table (>80% field-visit / 60–80% watch) with drivers & actions |
| Enterprise detail | 24-month history + fan chart with flag marker; UPI / rainfall / mandi / SHG sparklines |
| Liquidity calendar | Enterprise × month P(deficit) heatmap, cohort filter, click-through |
ml/pipeline.py panel simulation + quantile forecasting + JSON export (seed-locked to Round 1 deck)
frontend/ React + Vite + Tailwind + Recharts dashboard
docs/ 3-minute demo video script
code/, figs/ original Round 1 scripts and figures (kept for provenance)
All data is calibrated synthetic (kharif calendar, monsoon input costs, festival demand, drought regimes — every page carries the disclosure footer). The pipeline is signal-agnostic: Round 3 swaps the simulated blocks for live AGMARKNET, IMD and SHG-ledger feeds without touching the model layer.