Mandi information, in context — observed prices, market comparisons, reporting quality, and carefully explained seven-day outlooks across South India.
Live application · Explore markets · Compare markets · Methodology · Latest release
MandiLens turns official but irregular AGMARKNET reports into a focused public decision-support product. It helps farmers, Farmer Producer Organizations, traders, analysts, agritech teams, and market researchers inspect what was reported, compare like-for-like market dates, and understand how fresh and reliable each result is.
The product deliberately keeps observed data, estimates, and user-entered assumptions separate. It is information only—not trading, procurement, or financial advice.
- Search 189 represented market-commodity series across 120 markets.
- Inspect observed minimum, representative, and maximum prices with historical context.
- Explore reported arrivals, seasonality, missing-report periods, and anomaly flags.
- Compare two to four markets for one commodity on a common observed or forecast date.
- Add quantity and market-specific cost assumptions without inventing hidden costs.
- Review a lead-specific seven-day outlook with explicit uncertainty and reliability context.
- Read detailed reports covering data quality, methodology, sources, limitations, and architecture.
- Use the interface in English and nine Indian languages.
The deployed snapshot represents source data through 20 July 2026.
| Measure | Published scope |
|---|---|
| States | 6 |
| Districts | 78 |
| Markets | 120 |
| Commodities | 14 |
| Market-commodity series | 189 |
| Prepared market-day observations | 91,737 |
| Source variety rows inspected | 1,569,829 |
| Accepted variety rows | 1,564,481 |
Represented states are Andhra Pradesh, Karnataka, Kerala, Maharashtra, Tamil Nadu, and Telangana. The 14 commodities include vegetables, grains, oilseeds, fibre crops, spices, and plantation crops.
These captures show the current production interface on desktop and mobile.
Market Explorer
|
Guided market comparison
|
Market history and evidence
|
Responsive mobile experience
|
The selected forecasting method is a validated recent-level and lead-aware blend. Candidate methods are compared using chronological selection folds and a separate locked 60-day holdout, so future observations cannot leak into model selection.
| Locked-holdout measure | Result |
|---|---|
| Forecast examples | 195,253 |
| Mean absolute error | ₹521.20 per quintal |
| WAPE | 10.53% |
| Directional accuracy | 49.15% |
| Empirical interval coverage | 71.59% against an 80% target |
That coverage shortfall is published rather than hidden. Missing market reports also remain missing—they are never converted into a zero price. Read the live forecast reliability and data-quality reports for the full interpretation.
flowchart LR
A["AGMARKNET 2.0 public API"] --> B["Monthly cache and checksums"]
B --> C["Python validation and aggregation"]
C --> D["Parquet analytical layer"]
D --> E["Chronological model evaluation"]
E --> F["Forecasts and empirical intervals"]
F --> G["Compact static JSON"]
G --> H["Next.js application on Vercel"]
U["User quantity and cost assumptions"] --> I["Browser-only comparison"]
H --> I
Forecasts are produced by a reproducible batch pipeline and published as versioned artifacts. The web application does not need a request-time model server, database, login, or browser secret.
| Layer | Main tools |
|---|---|
| Web product | Next.js 16, React 19, TypeScript |
| Charts and interface | Recharts, Lucide, responsive CSS design system |
| Data pipeline | Python 3.12, Polars, Parquet, Pydantic |
| Forecasting | scikit-learn, rolling-origin evaluation, empirical residual intervals |
| Exploration | DuckDB |
| Quality | Pytest, Ruff, strict mypy, Vitest, ESLint, Prettier |
| Automation and hosting | GitHub Actions, Vercel |
- Node.js 24
- Python 3.12 or newer
- uv
make install
make devOpen http://localhost:3000.
Equivalent direct commands:
uv sync --all-groups --frozen
npm ci --prefix apps/web
npm run dev --prefix apps/web| Command | Purpose |
|---|---|
make check |
Run formatting, linting, type checks, tests, and the production build |
make download |
Retrieve and cache configured official source partitions |
make validate |
Validate source rows and rebuild quality evidence |
make transform |
Rebuild processed and published market-day data |
make train |
Run chronological evaluation and fit the selected method |
make forecast |
Publish lead-specific forecasts and intervals |
make refresh |
Refresh recent source months and rebuild the complete product artifact |
The source backfill is rate-limited, cached, and restartable. A failed refresh does not replace the last known-good published snapshot.
apps/web/ Next.js application and static web artifacts
config/ Pipeline, quality, model, and path configuration
data/published/ Versioned analytical datasets and manifests
models/published/ Selected model bundle and metadata
pipelines/src/ Ingestion-to-export Python package
reports/ Data-quality and model-evaluation reports
tests/ Pipeline, leakage, calculation, and artifact tests
docs/ Architecture, methodology, cards, and screenshots
.github/workflows/ Quality gates and scheduled data refresh
Thoughtful contributions are welcome. Good starting points include documentation improvements, accessibility fixes, clearer chart explanations, test coverage, and well-scoped data-quality bugs.
Read CONTRIBUTING.md before opening a pull request. For larger product, data, or model changes, open an issue first so the scope and evidence requirements are clear.
Open an issue · View the changelog · Browse releases
Security concerns should follow the private reporting guidance in docs/SECURITY.md.
Source data comes from the Directorate of Marketing and Inspection, Ministry of Agriculture and Farmers Welfare, Government of India:
Source attribution does not imply government endorsement. Market reporting can be delayed, incomplete, revised, or inconsistent across locations. Historical performance and forecast intervals are not guarantees for a particular market or date.
This repository welcomes issues and pull requests, but a software license has not yet been
selected. Public visibility does not itself grant software reuse rights. Add an explicit LICENSE
before presenting the code as formally open source.