Methodology & Validation
How we test and improve alert criteria using real data. This document is public because we believe transparency is part of safety.
AlertaParques uses weather data from Open-Meteo (ERA5/ECMWF model) to assess conditions on 73 trails in Chilean national parks. Like all numerical weather models, Open-Meteo has known limitations in complex mountain terrain. This page documents those limitations, how we detected them, and what we did to compensate.
Backtesting against real incidents
To validate our thresholds we queried the Open-Meteo historical API (ERA5) at the exact coordinates of 12 documented incidents between 2023 and 2026 — CONAF preventive closures, accidents, and fatalities in Chilean national parks. Below are the three most detailed cases, followed by a summary table of the nine additional ones.
Circuito O — Torres del Paine
17 November 2025 · Paso John Gardner · 1,241 m · Austral Zone
A blizzard with gusts up to 193 km/h and zero visibility caused the deaths of five foreign tourists. Open-Meteo, anchored at 244 m (997 m below the actual pass), recorded a maximum gust of 108.7 km/h — 56% of the real wind.
With the red threshold of 80 km/h configured for Circuito O, the system would have triggered a RED alert from 05:00 AM, maintaining it for 16 consecutive hours. The group left camp between 05:00 and 07:00.
Volcán Llaima — Conguillío National Park
17 May 2026 · ~2,000 m · South Zone (La Araucanía)
A hiker died after falling 600 metres during the ascent. Rescue operations were suspended due to strong winds and zero visibility that prevented helicopter flight. Open-Meteo, anchored at 2,621 m (621 m above the accident point), reported a maximum gust of 41.8 km/h and minimum temperature of −8.1°C.
With the original criteria, the system would have shown GREEN all day. The helicopter couldn't fly; AlertaParques said there was no problem.
Volcán Villarrica — Villarrica National Park
30 May 2025 · La Pingüinera sector · ~2,100 m · South Zone
A 38-year-old hiker died after falling 500 metres. The rescue helicopter could not land due to strong winds and fog. Open-Meteo, anchored at 2,828 m (728 m above the accident), recorded a maximum gust of 43.6 km/h and minimum temperature of −14.6°C.
The system would have shown GREEN. The −14.6°C temperature with wind >25 km/h should have triggered yellow, but sustained modelled wind was insufficient due to the same structural problem.
Nine additional cases (ERA5 historical)
For each event, the Open-Meteo ERA5 API was queried at the approximate time of the incident, the current system criteria were applied, and the result was compared against the real level (CONAF closure or fatality).
| Date | Trail | Emitted | Real | Event |
|---|---|---|---|---|
| 21 Jul 2023 | Circuito O | GREEN | RED | CONAF closure due to snowfall. 30–80 cm accumulated. ERA5 underregistered gust (49 km/h vs threshold 80). |
| 3 Aug 2023 | Volcán Lonquimay | YELLOW | RED | Madeleine Segovia killed. Ice slope. ERA5: −9.5°C, gust 12 km/h. Terrain danger not captured by weather model. |
| 19 Jun 2024 | Volcán Villarrica | RED | RED | Claudio Moreno missing (found dead 15 Aug). ERA5: −10.4°C in winter → RED triggered by temperature. ✓ Correct alert. |
| 1 Aug 2024 | Circuito O | GREEN | RED | CONAF preventive closure. Real gusts 60–90 km/h. ERA5 recorded 26 km/h — severe underregistration in Patagonia. |
| 1 Aug 2024 | Volcán Villarrica | YELLOW | RED | PN Villarrica preventive closure. ERA5: −5.4°C, gust 54 km/h. Temperature criterion returned YELLOW before gust check. |
| 1 Aug 2024 | Sierra Nevada (Conguillío) | RED | RED | PN Conguillío preventive closure. ERA5: gust 80.3 km/h → volcano RED triggered. ✓ Correct alert. |
| 1 Aug 2025 | Circuito O | GREEN | RED | 10 protected areas closed. Real gusts up to 120 km/h. ERA5 recorded 22 km/h — critical underregistration. |
| 21 May 2026 | Volcán Osorno | RED | RED | Hiker killed in 150 m glacial crevasse. ERA5: gust 45.7 km/h → volcano RED. ✓ Correct alert. |
| 23 Mar 2026 | Circuito O | GREEN | RED | CONAF closure due to landslides. ERA5: gust 49 km/h (threshold 80). Systematic underregistration in Patagonia. |
Pattern detected: ERA5 systematically underestimates wind in the austral zone (Torres del Paine). In the 5 Circuito O cases, the model captured between 22% and 56% of real wind. This explains the majority of failures in this dataset — not a weakness of the evaluation algorithm.
Complementary layer: official CONAF closures
The backtesting revealed a structural pattern: ERA5 severely underestimates wind in the austral zone, and lowering thresholds to compensate would generate false positives across the rest of the country. The most effective solution is to add an authoritative data source that does not depend on the weather model.
Since July 2026, AlertaParques integrates the official CONAF RSS feed (conaf.cl/feed/) as an independent second source. The hourly cron parses the feed, detects trail closure articles via keywords, maps the mentioned region to the parks in that area, and stores the closure with a 48-hour TTL. When an active closure exists, a distinctive banner appears on the park page — with higher visual priority than weather alerts and linked to the official notice.
ERA5 model precision
33%
4 of 12 incidents. The weather algorithm correctly predicted the danger level from ERA5 data. This number does not change with the CONAF integration — it honestly reflects the model's real limitations.
Full system coverage
~67%
~8 of 12 incidents. Conservative estimate of what ERA5 + CONAF RSS would have detected combined. At least 4 of the ERA5 failures correspond to CONAF closures the RSS would have automatically captured.
Why these are distinct metrics and don't add to 33%
ERA5 precision measures whether the weather algorithm correctly predicts the danger level. System coverage measures whether any source in the system — model or official data — would have issued a warning. Merging them would artificially inflate the first figure, concealing the model's real limitations in austral terrain. The 33% is the honest indicator of the algorithm; the ~67% describes what a user would have experienced in the full system.
Limitations of the CONAF integration
The system detects closures published on the conaf.cl RSS within 48 hours of publication. It does not detect closures communicated only by phone, radio, or on-site signage. Accidents caused by terrain hazard without a formal closure (e.g. Lonquimay Aug 2023) are not capturable through this channel. Detection depends on article text: if CONAF publishes without mentioning the region or uses unexpected terminology, the region mapping may fail.
Structural limitation of Open-Meteo in high mountain terrain
The ERA5 model divides the planet into a grid of approximately 9 km cells. Within each cell, Open-Meteo anchors the elevation to the highest point of the represented terrain. In complex mountain terrain this produces two opposite types of error:
Grid anchored below the terrain
The model underestimates wind because it interpolates from a lower elevation. Open-Meteo captured ~56% of real wind in Torres del Paine. Conservative wind thresholds on Patagonian trails compensate for this underregistration.
Grid anchored above the terrain
The model reports wind at the summit or in free air, not on the slope where the trekker is. On volcanoes, this produces readings of 40 km/h while helicopters cannot fly. Modelled wind is structurally unreliable.
Implemented adjustments
As a result of the backtesting, all trails are classified into four terrain types with differentiated evaluation logic for each.
Standard trail 51 trails
Patagonia, forests, valleys, coastline. ERA5 grid with ~40–56% underregistration. Conservative wind thresholds to absorb the error. Visibility criterion active below 3,000 m.
Volcano ascent 9 trails
Villarrica, Llaima, Osorno, Lonquimay, Mocho and others. Modelled wind is unreliable. Temperature is the primary criterion: below −10°C in winter → RED; below −5°C → YELLOW. Wind thresholds reduced to 60% as a secondary signal.
Mid-range mountains 8 trails
El Morado, Cerro Castillo, La Campana and similar (1,500–3,000 m). Grid reasonably reliable. 75% correction factor applied to wind thresholds.
Altiplano 5 trails
Lauca, Llullaillaco, Tres Cruces (>3,500 m). Primary temperature criteria. Seasonal logic for altiplano winter (November–March). 70% correction factor on wind thresholds.
Verified accuracy
Every evaluation generated by the cron is logged. When we can compare the emitted level against what actually happened on the ground, we record that verification. The current set of 12 verifications is the backtesting documented on this page. These numbers reflect the ERA5 model in isolation — without the CONAF layer. The global accuracy honestly reflects known limitations: ERA5 underestimates wind in Patagonia (the 5 Circuito O cases account for most failures) and does not capture terrain state on glaciers or accumulated ice. The "Complementary layer" section above describes how the CONAF RSS integration raises estimated system-wide coverage to ~67%.
Are you a park ranger, mountaineer, or outdoor professional?
If you visited a park and real conditions didn't match the alert — in either direction — that comparison helps improve the system. Write to us at alertaparques@gmail.com if you'd like to participate as a field validator.
External validation pending
The current thresholds are the result of the backtesting described on this page and international technical references (UIAA, MWIS, MeteoAlarm). We are actively seeking collaboration with meteorologists specialised in Chilean mountain weather to refine them by geographic zone. This page will be updated when that validation is complete.
Criteria last updated: June 2026 · View detailed thresholds →
How to use this information
AlertaParques is a support tool, not an official safety system. A GREEN level does not guarantee safety — especially on volcanic terrain where model limitations are greater. When in doubt, always check official communications from CONAF and SERNAGEOMIN, and carry appropriate equipment regardless of the system's status.