vnpredict
vnpredict.tech ↗Vietnam’s economy, month by month: 28 indicators tracked, 17 forecast and graded against a random walk.
Figures read from vnpredict.tech on 15 October 2026.
The problem
Traders, investors, and policy makers act on Vietnam’s monthly macro releases, which arrive from several official and market sources. A forecast helps them when it states its uncertainty and shows how often it has beaten the simplest alternative: carrying the last value forward.
The approach
- Fourteen series (growth rates, interest rates and FX reserves) blend a local AR(1) model with the Chronos-2 foundation model, 12 months ahead: Chronos-2 carries 70% of the weight at the first step and 20% at the last, because these series revert to their mean. The trade balance and the VN-Index use Chronos-2 alone.
- USD/VND is a managed peg, so it uses a robust-drift estimator (a Huber M-estimator of trailing monthly log-returns) with a conformally calibrated band; out of sample it beats both the foundation model and a flat-carry random walk, at MASE 0.88.
- Every forecast is graded over 8 to 12 rolling-origin cross-validation windows against a horizon-matched random walk: good below MASE 1, moderate below 1.5, poor at 1.5 and above.
- The other 16 forecasts use 80% bands from the model that produces them; the scorecard publishes how often backtest outcomes fell inside them (56% to 83%).
- A web-grounded LLM layer writes per-indicator and cross-indicator commentary at forecast time, refreshed with the data on a 14-day cycle.
Architecture
TradingView feeds + official releases → AR(1) and Chronos-2 blend, Chronos-2, robust drift (USD/VND) → 80% bands → rolling-origin grading vs a random walk → web-grounded commentary → dashboard and monthly report
Results
28 indicators tracked and 17 forecast 12 months out, each with a good / moderate / poor grade from 8 to 12 cross-validation windows. On 8 October 2026 the growth-rate and rate series moved from Chronos-2 alone to the AR(1) blend, after the blend beat Chronos-2 alone on the same cross-validation windows. The scorecard read 12 good and 5 moderate on 1 October, and 14 good and 3 moderate (MASE 1.04 to 1.09) on 15 October, with none poor; all 17 grades are published. Each issue of Vietnam Macro Monthly prints to PDF or downloads as an editable deck.
- Time-series foundation models
- Huber M-estimation
- Conformal prediction
- Rolling-origin cross-validation
- MASE vs. random walk
- Search-grounded commentary