What an avalanche report teaches you about deciding with imperfect data

I ski tour in the Alps most winters, and most tour mornings start the same way: standing in a cold car park before sunrise, coffee going lukewarm, phone out, reading the day's avalanche bulletin before anyone clips into a ski. The bulletin gives you a number, 1 to 5. It is tempting to read that number, decide, and go. Experienced tourers don't, because the number by itself is the least useful part of the report. What actually decides the day is everything underneath it, and that stack of underneath-the-number thinking turns out to be a better lesson in reading imperfect data than most dashboards ever teach.

What does the danger number actually mean?

The European Avalanche Danger Scale runs 1 to 5: low, moderate, considerable, high, very high. It has been the shared standard across France, Italy, Switzerland, Austria and Germany since 1993 (EAWS). Forecasters don't read that number off a sensor. It comes out of a consensus process, the EAWS matrix, that combines three separate judgements: how unstable the snowpack is, how widespread that instability is across the terrain, and how large a resulting avalanche could be (EAWS matrix study, 2025). Three fuzzy inputs, turned into one clean-looking digit. That compression is exactly what a KPI on a dashboard does too, and it is exactly where the trouble starts.

Why can two "considerable" days feel completely different?

Because the number hides which kind of problem produced it, and the problem types don't carry the same certainty. A wind slab, snow that piled up behind a ridge overnight, is visible, testable and moves predictably; a forecaster can point at the slope and tell you where it is. A deep persistent slab, a weak layer buried under months of snow, is nearly impossible to test directly, can sit dormant for weeks, and produces the avalanches that kill experienced people who read the bulletin correctly (problem definitions per Colorado Avalanche Information Center, a structure Europe's forecasters use the same way). Two "3, considerable" ratings, one built on a wind slab and one on a deep persistent slab, are not the same day. One number, two very different amounts of forecaster certainty behind it. A revenue dashboard that says "3.2%" behaves the same way: the number never says whether it came from a clean, complete pipeline or a spreadsheet somebody patched together at 11pm.

Why doesn't the bulletin just publish a confidence score?

Some do, in words if not always in a number. The Canadian avalanche industry formalised this as a separate confidence rating, defined as the forecaster's own degree of certainty about the prediction, and low confidence usually means the snowpack is changing fast or the observation network is thin that day (definition via Joe Stock's summary of the InfoEx standard). European bulletins more often fold that same judgement into the written discussion above the number rather than a standalone score, but the information is the same: how sure is the person publishing this. Most business dashboards skip both the score and the words. A single traffic light or one decimal number gets shipped with total confidence attached by default, whether the underlying pipeline is solid or held together with manual exports. That's the habit worth stealing directly: attach an honest confidence signal to a number, not just the number.

What does the bulletin show that a single number can't?

Location. Every European bulletin publishes danger broken out by aspect, which direction a slope faces, and by elevation band, because the same storm loads a north-facing bowl and a south-facing ridge completely differently. The result is usually drawn as a rose, a stylised top-down view of a mountain colour-coded by compass direction and height. The headline number is really a ceiling; the rose is where the actual, useful detail lives. That is the same relationship a summary KPI has to the dimension breakdown underneath it: the top-line number tells you there's a number, the segment-by-segment view tells you where the number actually comes from and whether it's safe to trust everywhere.

So who actually decides, if the number can't?

You do, every time, with whatever the bulletin gives you. Forecasters never promise to remove the judgement call; the entire discipline of ski touring assumes you'll read the danger level, the problem type, the confidence and the terrain rose together, then still make a human call about which slope to ski, because nobody is coming to make it for you. That's not a design flaw in avalanche forecasting. It's the honest acknowledgement that the data will never be complete enough to fully replace judgement, so the job of the report is to make the judgement call as informed as it can be, not to pretend the call isn't needed. A good business dashboard should aim for exactly that, and most stop one step short: they hand over the number and imply the decision is now automatic, dropping the parts, the confidence, the breakdown, that would actually help someone decide well.

Where does the metaphor stop working?

A few places, honestly. Avalanche forecasters run an expensive, purpose built observation network, weather stations, snowpack pits dug by hand, a network of professional and volunteer field reports feeding one bulletin every morning. Most companies have nothing close to that scale of dedicated sensing for their own numbers, and building it is real work, not a checkbox. The stakes also don't compare; nobody dies from a dashboard, though a bad decision from a false-confident one can certainly cost real money. And once you commit to a slope you're mostly locked into the outcome for the next thirty seconds, while a business decision, wrongly made on bad data, can usually still be revisited tomorrow. That last difference cuts in favour of dashboards, not against them: there is even less excuse for shipping a number with no confidence attached when you can actually afford to get it wrong and fix it, unlike a skier on a slope.

Is this worth fixing on your own dashboards?

If your team makes real decisions off numbers that come from a pipeline nobody has stress-tested lately, the fix looks a lot like the bulletin: know which numbers rest on solid ground and which ones are closer to a deep persistent slab, quietly serious and hard to test. I wrote about the failure patterns that put a number on shaky ground in the five medallion mistakes I keep unpicking, and the honest, boil-it-to-one-number habit itself in the surf forecast platform I built in the open. Automation and DataOps is the service page for the actual mechanism, tests that catch a broken number before your team ever sees it, described on the automation and DataOps page.

Book a free intake call. Ask any question you like about your own numbers; you'll leave with an answer or a clear next step, no obligation either way.