Eight days of warning
before the fire.
Twenty weather stations. Seven years of NOAA public data. 1.1M readings, zero labels. A pattern every threshold alert system missed, sitting in the data for eight days before the Woolsey Fire started.
DisclosureBenchmark study on public NOAA.gov data. No customer data was used. Reviewed by Drew Wasem, Founder, Coherany. Methodology available on request.
THE METHODOLOGY
We looked at how signals moved together, not whether any one of them crossed a line.
On November 8, 2018, the Woolsey Fire ignited in the mountains above Los Angeles. Within hours it was racing toward Malibu. By the time it burned out, it had consumed 97,000 acres, destroyed 1,600 structures, and killed three people. Every fire weather alert system in Southern California was watching that day. The alerts fired. They were too late.
Threshold alert systems ask each station in isolation whether its current reading has crossed a fixed limit. That works for catastrophic spikes. It misses everything that builds slowly across a region, and fires build slowly across regions before they ignite.
We took the opposite approach. Instead of asking each station whether it was dry right now, we asked how the whole network was moving together over the last three days: which stations were drying, in which direction the drying was spreading, and which normally-independent signals had started moving in lockstep.
THE DATA
Seven years of the public weather record across the LA basin.
THE FINDINGS
Three signals that single-station thresholds cannot see
None of these patterns are visible to any alert system that watches stations one at a time. All three were building for days before the Woolsey Fire started.
Advance trajectory, not instantaneous value
A threshold system that saw 23% humidity on October 31 saw nothing unusual. A trajectory view that saw 53% → 35% → 23% over 72 hours saw a collapse in progress. The shape of the change matters more than the reading itself, and trajectory leads threshold by days.
Regional spread, not station isolation
Inland desert stations began drying 18 hours before coastal stations did. The cascade moved at the speed of the Santa Ana winds. Knowing the direction and speed of a drying event tells you where a fire is most likely to start, before it starts.
Coordinated movement, not independent readings
Stations that normally behave independently began moving in lockstep across zones that usually have nothing to do with each other. Correlated behavior in normally-independent signals is one of the most reliable early-warning patterns you can find in any operational dataset.
THE RESULT
The pattern had been sitting in the public data for eight days.
The approach surfaced the Woolsey pre-fire signal on October 30, more than a week before the fire started. It had never seen this fire. It had never been trained on fire data. It was discovering the pattern from the data itself and flagging the eight-day departure from normal as it happened.
Total time from raw data to discovered pattern: under an hour. Total labeled training data required: zero. Total cost of the data: free.
“The signal had been sitting in public data for over a week. No alert system is built to see it, because no alert system is built to look at how stations move together over time.”
HONEST LIMITS
What this does and doesn't prove.
This was a retrospective analysis on public NOAA data, not a live deployment inside a fire agency. A real operational system will need more stations, more validation against more fires, and integration with the agencies that actually dispatch crews. What the benchmark proves is that the signal is real and the architecture works on data that has been sitting in the public record for years. The only question is how much earlier we can get the warning to the people who can act on it.
This isn't really about wildfire.
It's about what's already sitting in your operational data that nobody has looked at the right way yet. Request the methodology or run it on your own data.