Select a market to see its forecast.
Price history and forecast
The shaded band is the 80% range: eight times in ten, the price should land inside it. The dashed line is what you would predict by assuming nothing changes.
Ask a what-if
These sliders change the model's inputs and re-run it here in your browser — no server, nothing sent anywhere. The trees are downloaded once, about 2 MB.
How well does this actually work?
Every forecast above is produced by a model that was tested the hard way: retrained at dozens of points in the past and scored only on months it had never seen.
Do the ranges hold up?
Measured on months held out of calibration. Closer to 80% is better.
Accuracy against the benchmark
Average error in taka across all test runs. Lower is better.
What this does not claim
Prices are the World Bank's modelled monthly estimates, not receipts from a till. Direct observations are sparse in earlier years.
Flagged anomalies are statistically unusual months. They are not evidence of hoarding, fraud or crisis.
Forecast skill depends on conditions. In calm months the benchmark is hard to beat, and the charts above say so rather than hiding it.
Market groupings describe price behaviour. A group is not a ranking, and its number carries no order.