Harvora™ model-card brief
Frost risk forecasting for New Zealand orchards model card
This model card documents the method, evidence target and limitations for frost risk forecasting for new zealand orchards. This is a held-out regional weather benchmark, not yet commercial orchard-outcome validation. The alarm threshold is deliberately tuned to miss almost nothing, so about half of flagged nights are true frost events (positive predictive value ≈ 46.91%) — the trade-off is published, not hidden. Full detail (TP 76 / FN 8 / FP 86 / TN 2,186, rounded from committed four-decimal rates) is in the technical brief.
Direct answer
What this supports
Each block gets a probability distribution of overnight minima and freeze chance — so a frost probability reads directly against your protection decision for that night, rather than a single alarm number.
Intended for New Zealand growers, technical partners, lenders and insurers.
Evidence and method
How the signal is formed
Hourly ensemble forecasts are bias-corrected per block and sampled as a 1,000-draw Monte Carlo distribution of overnight minima and freeze probability.
Evidence grade: ESTIMATE
Weather-benchmark measured today; every grower outcome logged in the product feeds the field benchmark it is measured against next.
Boundary
How to read this evidence
This is a held-out regional weather benchmark, not yet commercial orchard-outcome validation. The alarm threshold is deliberately tuned to miss almost nothing, so about half of flagged nights are true frost events (positive predictive value ≈ 46.91%) — the trade-off is published, not hidden. Full detail (TP 76 / FN 8 / FP 86 / TN 2,186, rounded from committed four-decimal rates) is in the technical brief.
Sources and related evidence
- Open-Meteo historical forecast API · SOURCED
- Harvora public technical brief · SOURCED
- Harvora source register · SOURCED
Published 25 August 2026 · Last reviewed 25 August 2026 · Harvora technical review
Apply it to your operation