von Eeva Latosuo , Pascal Haegeli

07. May 2025 - 12 min Lesezeit

Caution: Avalanche forecast may contain uncertainty

Users’ awareness and response to uncertainty information in public avalanche forecasts.

Uncertainty is an inherent part of avalanche risk management. Can you remember a single day without experiencing any uncertainty when planning for or travelling in the backcountry? Without any uncertainty, there would be no risk to manage. In addition, every day is different. On some days, there are only a few unknowns, and on other days, the whole day is stacked with more questions than answers.

Even though assessing avalanche conditions is public avalanche forecasters’ full-time job, they do not always know what is going on. They do not know what is happening in every part of their forecast region or they do not know the exact timing of the storm that is predicted to dump a significant load on the snowpack. However, there is no consistent way how uncertainty is communicated in avalanche forecasts.

Talk uncertainty?

Some forecast centers do talk about uncertainty. For example, Avalanche Canada uses confidence statements at the very end of their forecast text that give a general perspective to the unknowns of the day without providing the reader much tangible advice.

The US avalanche centers do not have a specific placeholder for discussing uncertainty, but forecasts, at least sometimes, talk about this conundrum. Last March, a forecaster wrote in the bottom line: “The new snow may be sensitive. On the other hand, it may be stable.” This is a lot of words for simply saying “Sorry, I don’t know.” It seems that we want to communicate about uncertainty since it is a critical component of our assessments, but we do not know how to do it well. So, how can uncertainty be communicated more meaningfully in public avalanche forecasts?

Exploring the Alaska Range. Photo: Eeva Latosuo

To address this challenge, we have embarked on a multi-year research project that aims to develop and evaluate useful and feasible ways to communicate the extent and characteristics of uncertainty in public avalanche forecasts. This is a collaboration between the Colorado Avalanche Information Center, National Center for Atmospheric Research, and Simon Fraser University. We will combine three meaningful perspectives:

We want to learn from best practices in related fields like meteorology,
We want to hear from forecasters about how they conceptualize uncertainty and what they think is feasible to communicate, and
We want to learn from users what they think is actually helpful for their decision-making process under uncertainty.

In this article we are sharing the initial results of what we have learned from the users of avalanche forecasts in an online survey.

Uncertainty in avalanche forecast products

Let’s first look at a general framework for uncertainty in avalanche forecasts. One source for unknowns is natural variability or inherent randomness in the natural processes that lead to avalanches. The technical term for this type of uncertainty is aleatoric uncertainty. A coin toss and the roll of a dice are good examples of this type of uncertainty. Even though the process is quite simple, we cannot confidently predict the outcome due to factors that cannot be controlled. An important quality of this type of uncertainty is that it is impossible to reduce it no matter how hard we try to understand what is going on.

Another source of uncertainty is system complexity. Complex systems have so many interactions and feedback loops at different scales that it is difficult to understand them in their entirety and hence make accurate predictions. This type of uncertainty is called epistemic or knowledge uncertainty, and a good example of it is weather, which is difficult to predict accurately even if the current weather models have reduced uncertainties through increased knowledge and innovative research.

The third type of uncertainty that we consider is ambiguity, which is a property of the information we have about a system that allows for multiple interpretations. It often arises from vague definitions, incomplete observations, or conflicting evidence, where many interpretations seem reasonable given the available information.

We can apply these concepts to the production of avalanche forecasts (Figure 1). We have an avalanche phenomenon, which we are trying to capture with observations and models. This information is then interpreted by avalanche forecasters who describe their assessments in the forecast. Each of these steps introduce new sources of uncertainty which ultimately affect the decision-makers who use the information for deciding when and where to travel in the backcountry.

Figure 1: Framework for the sources of uncertainty in avalanche forecast products.

The inherent randomness and complexity of the avalanche system results in spatial and temporal variability that makes predicting individual avalanches with certainty almost impossible. We gather lots of observations and use computer models to try to capture what is going on, but they provide incomplete evidence and have their own challenges that introduce additional uncertainties. The interpretation of these data by forecasters are subjective judgments that are affected by their personal perceptions, knowledge and biases, but also by forecast procedures, like at what time a forecast must be published. Finally, the structure and limitations of the forecast products themselves introduce additional uncertainty into the forecast, including ambiguous or complex language. There are limits to what we can say with the text field and graphics we have currently available. In the end, the decision maker must deal with these uncertainties and combine them with their own knowledge and perception of uncertainty to come up with informed decisions about when and where to travel in the backcountry. It is surprising that despite its prevalence, we do not talk about uncertainty more explicitly in the forecast itself nor in avalanche safety courses.

Survey design

To gain first insights from forecast users, we conducted a detailed online survey in collaboration with the Colorado Avalanche Information Center last winter. The goal was to learn more about how users perceive uncertainty in the forecast, how much they know about it, and how they would respond to explicit uncertainty information. Our survey included several components, but its core consisted of hypothetical but realistic avalanche forecast scenarios. Each participant was presented first with one baseline forecast without any explicit uncertainty information. To examine how participants respond to uncertainty information, we modified these baseline forecasts by adding information about sources of uncertainty in the forecast discussion. The added uncertainty information related to the timing of the hazard, where one can find the avalanche problem in the terrain, and information about the quality of observations the forecast was based on. For example, we added sentences about high uncertainty around timing and location, and low uncertainty with respect to observations.

In some of the modified forecasts, we also added more synthesized information about uncertainty to the bottom line. In some we added a statement about the overall magnitude of uncertainty based on the information that was provided in the discussion (e.g., “We have a very low amount of uncertainty (Level 1 of 5) in the forecast today.”), and in some, we added travel and terrain advice that explicitly talked about how to deal with the uncertainty (e.g., “Conditions will be more manageable if you start and end your day early.”).

Figure 2: An example of a modified forecast used in the survey. This situation had a moderate danger rating with persistent slab avalanche problem. Colored text segments show how we embedded the explicit uncertainty statements in the forecast format. This example included overall magnitude statement (orange), travel guidance (yellow), high uncertainty about timing (green), high uncertainty about spatial distribution (blue), and low uncertainty about the observations (purple). Highlights and value codes are just shown to spotlight added statements. They were not shown to participants.

In total, we had 60 modified forecasts that all had different combinations of factors and different levels of uncertainty. Each participant was presented with one baseline and three modified forecasts, and for each we asked

How much uncertainty do you think the forecaster faced when producing this forecast? (0-100 rating)
Would you consider travelling in the backcountry under these conditions? (Yes/No)
How useful was the presented avalanche forecast for your decision? (0-100 rating)
How difficult did you find understanding the forecast? (0-100 rating)
How would uncertainty information like this impact your trust in the avalanche center as a source of reliable avalanche safety information? (7-step Likert scale from “much decreased trust” to “much increased trust”)

The scenarios were designed in a way that allowed us to isolate the effect of each of the modifications on the answers of our participants.

In total, we had 1313 survey participants completing all the relevant questions. Our sample was fairly experienced and committed to their winter backcountry activities. 75% were male, and the majority was 25–44-year-old. They were mostly university educated, and the majority had introductory level avalanche training.

User responses to uncertainty information

Let’s look at the main results. First, participants’ estimate of uncertainty increased substantially between the baseline forecasts and the modified forecasts. This means that simply talking about uncertainty in the forecast made people more aware of this quality of the avalanche hazard environment. Participants’ responses indicated that they were most responsive to differences in the uncertainty related to observations. Their answers also showed that they had some awareness that persistent avalanche problems have inherently more uncertainty than storm and wind slabs, but they were not aware that wet slab avalanche problems are also very difficult to forecast. They also responded to uncertainty in location and timing, but only about half as much than in the other factors.

Figure 3: Participants’ uncertainty estimates as a function of the number of uncertainty sources at High and the uncertainty magnitude statement being absent (blue) or present (red). Having the uncertainty magnitude statement improved participants’ understanding that uncertainties from different sources add up.
In those forecasts where we explicitly told them how uncertain the situation is using an explicit statement about the magnitude of the uncertainty; participants had a much better perception about the accumulation of overall uncertainty (*Figure 3*). Without the magnitude statement, uncertainty estimates rose with the presence of zero to two uncertainty sources at level High but levelled out with more sources (*Figure 3*: blue box plots). However, they continued to increase linearly with cumulative number of sources when the statement was present (*Figure 3*: red box plots).

We also learned that participants’ higher levels of perceived/estimated uncertainty are associated with more conservative choices about entering avalanche terrain. This shows that including uncertainty information impacts participants’ actions in the right direction. Participants’ responses showed that they have some awareness that conditions with higher levels of uncertainty require a wider margin of safety.

However, increased estimated uncertainty was also associated with lower ratings of usefulness of the forecast for personal decision-making. While initially a disappointing finding, it makes sense. Increased uncertainty makes decision-making about where and when to go in the backcountry harder, and if one had the choice between a product with more or less uncertainty, we would all choose the more certain product. But we do not have this choice in avalanche forecasts. There is generally only one product, and sometimes there is just a lot of uncertainty. Hence, this result seems less related to the shortcoming of the forecast, but more related to the challenge of how we make decisions with more unknowns.

On the other hand, adding uncertainty information did not make the forecast text more difficult to understand. So, participants did not find the uncertainty information difficult to understand but did not think it was useful. Our interpretation of this juxtaposition is that users will need support and guidance on how to meaningfully integrate uncertainty into their decision-making process. Avalanche warning services and avalanche education providers are in key positions to establish a common vocabulary for communicating uncertainty and to share practical suggestions for how to deal with it in the field.

Finally, adding uncertainty information increased our participants’ trust in the avalanche center as a reliable source of information. 70% of participants would trust the warning service more if it included uncertainty information in their forecasts like the survey scenarios. Trust in the avalanche center would decrease only for 5% of the participants. This was one of the strongest results we had in our survey.

Summary

In summary, there are two clear take-aways from our user research so far:

Including uncertainty information in the public avalanche forecasts has clear benefits with no real downsides.
Just adding uncertainty information is not enough, guidance on using uncertainty information effectively is crucial.

But the perspective of users is just one piece of the puzzle. To take this further, we are interviewing public avalanche forecasters at the Colorado Avalanche Information Center and Avalanche Canada this winter to better understand how they conceptualize uncertainty and how we could make the assessment of uncertainty more systematic and transparent.

This is all setting a stage for next season, when we want to design prototype products to communicate uncertainty in avalanche forecasts more effectively. We will do this in collaboration with both forecasters and forecast users to ensure the approach is both feasible and useful. Our goal is to offer recommendations about how to communicate uncertainty to different types of users since a single product might not work for everybody in the diverse group of people using public avalanche forecasts.

So, stay tuned for more insights—while uncertainty is unavoidable in avalanche forecasting, confusion about how to communicate about it doesn’t have to be.

Acknowledgments

We are grateful to the Coast Salish peoples on whose traditional and unceded territories Simon Fraser University and our research program resides. We thank the participants of the survey for their contributions to this research and Colorado Avalanche Information Center staff for the fundamental help with the outreach effort and the design of the forecast scenarios. Simon Trautman supported the outreach with the U.S. National Avalanche Center. Funding for the research was provided by the Colorado Avalanche Information Center.