All use cases
Use case · Time-series forecasting

Forecasting & planning

Forecast demand, revenue, prices, load and cash flow with time-series foundation models that generalise to series they have never seen, then fit them to your history. Every number comes with a range your team can plan against and a grade against a simple baseline.

28macro indicators tracked
14 / 17forecasts beat the random walk
0.88MASE on USD/VND, a managed peg

When you need this

You plan inventory, capacity, hedges or budgets on single-number forecasts, and nobody can say how wrong they are. Vendor models arrive without a benchmark, so you can’t tell skill from luck. Seasons shift, markets turn, and last year’s model goes stale while the decisions keep flowing.

What we build

  • Forecasts that prove their value: time-series foundation models, cross-validated against simple baselines (random walk, seasonal naïve, drift). Each forecast carries its grade, and the misses are published with the hits.
  • Ranges you can plan against: each 80% band is checked on rolling backtests against how often the outcome landed inside it, and that coverage is reported beside the forecast.
  • Scenarios across horizons, so planners see the spread of outcomes.
  • A closed loop: each forecast is scored against the actual when it arrives, the model learns from the miss, and an update goes live only if it scores no worse than the current model on your metric.
  • Written analysis, optional: a web-grounded language model turns each forecast release into notes your team reads beside the chart.

How it’s built

Your data pipeline → time-series foundation models + robust estimators → conformal calibration → REST API · dashboards · alerts

Methods: time-series foundation models · classical quantitative models

In our products

vnpredict tracks 28 indicators of Vietnam’s economy and forecasts 17 of them 12 months ahead. Each forecast is graded over 8 to 12 rolling-origin cross-validation windows against a horizon-matched random walk: on 15 October 2026, 14 of the 17 beat it (MASE below 1), and the other 3 are published as moderate. Fourteen forecasts blend a local AR(1) model with the Chronos-2 foundation model, two use Chronos-2 alone, and USD/VND, a managed peg, uses a robust-drift estimator (a Huber M-estimator) with a conformally calibrated band that scores MASE 0.88. The blend went live on 8 October after it beat Chronos-2 alone on the same tests, and lifted the count that beats the random walk from 12 to 14.