Many natural hazards are hard to predict on time scales useful for risk reduction. This arises both from incomplete knowledge of underlying processes and from intrinsic unpredictability. The latter stems from strong nonlinearities (e.g. chaos), fine-scale aggregation in space and time, high dimensionality, and the open nature of Earth systems that exchange energy and matter with their surroundings in often uncontrolled ways. As a result, the future evolution of a natural hazard can only be expressed probabilistically through forecasts: each forecasting model provides a probability distribution for hazard intensity within a given space–time window. Even predictions, including point estimates (“best guesses”), implicitly carry uncertainty, whether stated or not, and are therefore effectively probabilistic. In addition to intrinsic variability, pervasive knowledge gaps often require using multiple forecasting models, introducing further probabilistic complexity in the Earth sciences (Oreskes et al., 1994) and complicating effective communication.

These issues have been extensively discussed within the RETURN project (multi-Risk sciEnce for resilienT commUnities undeR a chaNging climate; https://www.fondazionereturn.it/en/, last access: 5 October 2026). RETURN is a large national partnership that brings together eighteen Italian universities and research centers, the Italian Civil Protection Department, Italian State Railways, Assicurazioni Generali, other profit entities, and one Italian River Basin Authority. It has been designed to strengthen Italy's research capacity on environmental, natural, and human-induced risks, while connecting it with major European and global value chains. The project had a dual mission: on the one hand, to advance basic knowledge and develop new technologies for risk prevention and mitigation; on the other, to transfer these innovations into practical applications, engaging public administrations, companies, and local communities. Within the RETURN framework, we established a transversal task force comprising scientists from multiple disciplines, communication experts, and decision-makers. Its aim was to address the problem of handling uncertainties in hazard forecasting from the perspectives of scientists, practitioners, communicators, and stakeholders across a range of natural hazards – including floods, earthquakes, volcanic phenomena, landslides, climate projections (which can be interpreted as forecasts conditioned on a specific emissions scenario), weather forecasting, and marine biogeochemistry. Although this paper focuses on hazard forecasting, the proposed framework represents an important first step toward a comprehensive treatment of uncertainty in risk forecasting. In its simplest formulation, risk is a function of hazard, exposure, and vulnerability. While uncertainty in hazard has received considerable attention, vulnerability assessments are also affected by substantial uncertainty, which may in some cases exceed that associated with the hazard itself (e.g. Pianosi et al., 2026). We argue that the probabilistic framework proposed here for representing and propagating uncertainty in hazard forecasting can be extended naturally to the vulnerability domain. If this is the case, the resulting risk assessment becomes the combination of the uncertainties associated with both hazard and vulnerability, propagated through the risk model.

Drawing on the experience as both producers and users of such forecasts, we summarize here the main concepts discussed within the task force. First, we outline the general discussion and highlight common challenges across different hazards that may warrant shared procedures (e.g. Beven et al., 2018a). Then, we examine these scientific and communication challenges in detail. Finally, we offer general recommendations to help scientists and practitioners to develop common procedures for building and testing forecasts, clarifying the meaning of a “complete forecast”, i.e. a forecast that explicitly represents and keeps distinct all recognized uncertainties, and communicating these uncertainties.

The first issue that clearly emerged from the discussion inside the task force is the lack of a common terminology across the different disciplines as well as the different risk management stakeholders. For example, the terms prediction and forecast are often used with different meanings across disciplines. In this paper, a prediction is defined as a deterministic statement that a specific hazard intensity will or will not occur in a particular geographic region and time horizon, whereas a forecast gives a probability that such an event will occur (Jordan et al., 2011). Although also predictions are inherently probabilistic, due to the unavoidable presence of false alarms and missed events, we argue that issuing a forecast is preferable because it quantifies uncertainties through probability, and it clarifies the roles of the different actors in the whole decision-making process: in a nutshell, scientists provide probabilities describing scientific uncertainty, whereas decision-makers select probability thresholds to guide actions (Jordan et al., 2014). Section  summarizes the terms used in this paper and their meanings as commonly applied across different natural hazards. It does not claim to provide the “correct” definitions, but it is intended to facilitate readability and understanding.

We also noticed that different hazards share common features and challenges: (i) the presence of uncertainties linked both to the natural variability of the process and to our limited knowledge; (ii) the lack of a clear and unambiguous definition of these uncertainties and how to include both of them in a complete hazard forecast; (iii) how to test the reliability of the complete forecasts; and (iv) how to communicate these forecasts effectively to decision-makers and society in general. Oddly, despite the many similarities across fields, these problems have largely been tackled in isolation, as evidenced, for example, by the proliferation of distinct terminologies used to describe and to handle different kinds of uncertainty: without pretending to be exhaustive, terms like shallow and deep, intra- and inter-model, external and internal, value and structural uncertainty, likelihood and confidence, state and model uncertainty, uncertainty on model parameters and on initial/boundary conditions are widely used across many disciplines to deal with similar issues, both within and beyond natural-hazard forecasting (e.g. Marinacci, 2015).

These common features and distinctive characteristics are summarized in Fig. 1, which illustrates the different operational perspectives according to the role of the task force, regardless of the specific hazard considered. The first block defines the operational context, i.e. whether the forecast is long-term (e.g. forecast over years to tens of years or centuries, guiding structural design and planning) or short-term (e.g. forecast over hours/days/months, managing an unfolding emergency).

The second block covers the construction of forecasting models selection according to the context, i.e. the time horizon investigated (short- or long-term) and scenario projections (like for IPCC). Depending on the hazard, forecasting models can be of different types, for example: (i) Deterministic physics-based models, often accelerated through surrogate or reduced-order models (de Burgh-Day and Leeuwenburg, 2023; Li et al., 2023), incorporate uncertainty into their point-forecasts either by introducing random perturbations to the initial and/or boundary conditions or by augmenting deterministic point predictions with probability distributions calibrated from historical discrepancies between model predictions and observations (Glahn and Lowry, 1972; Gneiting et al., 2005; Koutsoyiannis and Montanari, 2022; Shabestanipour et al., 2023); (ii) Artificial Intelligence/Machine-Learning models that yield predictions with associated predictive errors (e.g. Mosavi et al., 2018; Reichstein et al., 2019; Mondini et al., 2023;); (iii) empirical or stochastic models derived from instrumental and/or historical data (Ogata, 1988; Gumbel, 1958; Stedinger and Cohn, 1986; Merz and Blöschl, 2008); (iv) models based on expert judgment (O'Hagan, 1998; Aspinall and Cooke, 2013). This list shows some end-members but mixed approaches are common; for example, data-driven Bayesian approaches may use different pieces of information (e.g. from data, physics, and expert judgement) to produce a combined forecast (e.g. Viglione et al., 2013), or physics-based model to produce a starting forecast that is continuously updated as soon as new data and information becomes available (e.g. Bluecat; Koutsoyiannis and Montanari, 2022; Montanari and Koutsoyiannis, 2025). Regardless of type, each model provides one probability distribution for the hazard (or risk) intensity of interest. The presence of single vs. multiple distributions strongly affects the scientific evaluation of the forecasts, communication, and use. For example, a single probability distribution yields a single probability for a specific event and implies a very good knowledge about the process and the underlying modelling procedure. Multiple distributions imply that the probability of that event is itself described by a set of probabilities rather than a single value, underlining our limited knowledge about the process.

The third block addresses forecast scientific evaluation perspective, which is essential for establishing the operational credibility of forecasts for practical use. The evaluation procedures depend critically on two key factors: the availability of data for testing and the independence of those data from the model that generates the forecasts. These factors dictate both the appropriate testing procedures and how to interpret the results. This last aspect will be deeply investigated in Sect. 5 “Evaluating the reliability and skill of forecasts”.

In Sects. 3–6, we describe all these common challenges in greater depth.

Uncertainty arises from multiple sources, which can be grouped into two broad categories in forecasting: the natural variability of the data-generating process and our limited knowledge of the true data-generating process or the limitations of its numerical representation. These are commonly referred to as aleatory variability and epistemic uncertainty, respectively (see Sect. ). Here, we implicitly assume the existence of a true data-generating process, an assumption to which we return later in the paper.

As implied with Fig. 1, regardless of the adopted modeling approach to describe the hazard intensity of interest, the model output is a forecast intended to describe the natural variability of the process. One common challenge across different hazards stems from the existence of multiple models producing forecasts or projections. In many current practices, these distinct forecasts are aggregated into a single forecasting distribution for the hazard intensity of interest (e.g. Rougier and Beven, 2013; Beven et al., 2018a). This forecasting distribution may be constructed as a mixture of the individual distributions (although, in some fields, this is incorrectly referred to as the “mean distribution” rather than the mixture distribution) or by using other statistically sound combination methods (Gneiting and Ranjan, 2013). However, this approach is hardly satisfactory for several reasons, the most obvious one being that it does not convey an important message to the stakeholders, i.e. how much scientists believe in their final assessment. For example, stating that a hazardous event has a 15 % probability of occurring conveys nothing about whether that estimate comes from a large or small dataset, whether forecasting models agree, or whether it is merely the average of widely divergent assessments.

In fields such as climatology and seismic hazard, practitioners often separate different types of uncertainty to indicate the faith they place in their forecasts. Specifically, climate change projections and long-term seismic hazard analysis acknowledge the existence of these different uncertainties (IPCC 2021; Senior Seismic Hazard Analysis Committee (SSHAC), 1997; see also discussion at Sect. 4), even though this is done through approaches and procedures that are incommensurable (Aven and Renn, 2015). Although all of these approaches are undoubtedly important steps toward more fully incorporating and representing uncertainties in a complete forecast, many critical issues of different types remain unresolved.

Every forecasting model is, by definition, a scientific product and must be tested against independent data (Box, 1976; AAAS, 1989; Klemeš, 1986; Schorlemmer et al., 2018). We contend that a forecast operational credibility is closely tied to the scientific reliability of the model, i.e. the model's ability to produce forecasting distribution(s) that adequately describe independent observations (Merz et al., 2024). Put differently, for societal applications the scientific reliability of a forecasting model is a necessary (sine qua non) condition for judging its credibility, before any evaluation of its effective applicability. Here, we do not dwell on specific test types (those depend heavily on the quantity and quality of available data and forecasts); rather, we focus on defining a general testing framework that can be shared across a broad set of practitioners in diverse scientific and applied domains characterized by varying time horizons, employing different model types and, crucially, facing widely varying amounts of data for testing.

Following conventions in other fields (e.g. Gneiting and Katzfuss, 2014; see also Schorlemmer et al., 2018 for seismology applications), the scientific evaluation of forecasts using statistical tests rests on two independent pillars: (1) comparing a FD with observations, and (2) assessing the relative performance (skill) of FDs generated by different models. To correctly interpret their outcomes, we clarify these two pillars and the adopted terminology in the following.

The first pillar concerns the reliability of a FD, or of a complete forecast. Broadly speaking, they are reliable when it agrees with observations. More specifically, we use the term “calibration” of a single FD (see Sect. ) when comparing it with observations that are independent from the model that produces the FD (Laio and Tamea, 2007; Gneiting and Katzfuss, 2014; Schorlemmer et al., 2018). Note that calibration has different meanings across communities; for example, in some fields it denotes the process of estimating model parameters; while perfect terminology may not exist, we agreed on the need for clear definitions accessible to all users, which motivates the glossary appended to this paper (Sect. ). When comparing the complete forecast (see the example of Fig. 5), which explicitly represents and keeps distinct aleatory variability and epistemic uncertainty, with independent observations, we use the term “validation” of the complete forecast (Marzocchi and Jordan, 2014).

If the data are not independent of the model's construction, we use the less ambitious term consistency. In many cases, observations are typically scarce and often not independent. Consequently, FDs are seldom fully validated or calibrated in the strict sense; instead, FDs are tested for their consistency with available observations. This distinction is substantive: for example, a flawed FD will typically fail a calibration test, but it can still pass a consistency test by overfitting the data used in its construction. In such cases, only the rejection of a consistency test is informative, as it reveals possible ontological errors or “unknown unknowns”.

Testing the calibration of a single FD and the validation of a complete forecast have profound conceptual and technical differences. When a single FD is available and independent observational data exist, the calibration compares the FD with the empirical distribution of the data. If the model is calibrated, the two distributions have to be similar. Many statistical tests exist for this purpose; a useful review is provided by Gneiting and Katzfuss (2014). One widely used technique to test a single FD is the Probability Integral Transform (PIT), a type of calibration plot: for each observed value within a forecast interval Δt, it computes the cumulative probability of that observation under the forecast distribution for Δt. For a calibrated forecast, the PIT values should be uniformly distributed (e.g. Laio and Tamea, 2007).

The validation of a complete forecast follows a different logic (see, e.g. Marzocchi and Jordan, 2014). Specifically, it consists of formulating the ontological null hypothesis, ϕ^∼p(ϕ). This null hypothesis can be tested in many different ways, depending on the available data and the context (see, e.g. Marzocchi and Jordan, 2018; Gerstenberger et al., 2024; Pauli and Parolai, 2026). Pragmatically, we check if the observed frequency of exceedances for a given hazard intensity x0 is coherent with the EED, p(ϕ). The distribution p(ϕ) is built fitting the set of ϕm with a Beta distribution, though any suitably chosen distribution could be used. The choice itself is a potential source of ontological error; omitting this uncertainty is equivalent to assuming a Dirac (point-mass) distribution, which is a far stronger, and more questionable, assumption.

The validation of a complete forecast becomes more challenging in the presence of non-stationarity. In this case, it is generally impossible to estimate an empirical exceedance frequency for a given hazard threshold x0 at a specific time because only one (or very few) observations are available for each time instant. However, if the non-stationary model is correctly specified, the exceedance probabilities associated with the observations should be uniformly distributed, according to the probability integral transform (PIT) (Gneiting and Katzfuss, 2014). Although this problem has not yet been fully addressed in the literature, the PIT framework can, in principle, be generalized to account for epistemic uncertainty. Let fm(⋅) denote the FD associated with the mth model and let x(ti) be the observation at time ti. For each FD, we compute the PIT values ϕm(ti)=fm(x(ti)). For the true FD (i.e. the distribution associated with the data-generating process), the corresponding set of PIT values {ϕm(ti)} should follow a uniform distribution. When epistemic uncertainty is represented by an ensemble of M alternative FDs (m=1,…,M), the ontological null hypothesis can be explored by requiring that the ensemble of PIT distributions adequately encompasses the uniform distribution.

The second pillar addresses relative forecasting skill across multiple FDs. The (relative) skill is measured by proper scoring rules that rank models by their ability to explain a given set of observations, without implying reliability (Gneiting and Katzfuss, 2014). The distinction matters: among competing FDs, one FD will often fit the observations better, yet all FDs might still be unreliable (i.e. fail reliability tests). Conversely, several FDs may pass reliability tests but differ substantially in skill, e.g. one may explain observations much better than the others. When observations are independent of model construction, higher skill indicates a better FD. If the same observations were used to build the FDs, superior skill may simply reflect overfitting.

In Fig. 7 we summarize the whole evaluation process and the different alternatives that depend on the availability and amount of data, the kind of test (reliability of a forecast or comparison of a set of forecasts), and the inclusion, or not, of the epistemic uncertainty in the reliability tests using dependent and independent data.

Needless to say, the least desirable situation arises when few or no observational data are available. For example, a model may be testable only over a limited range of hazard intensities, while observations of the highest, and often societally most relevant, hazard intensities are unavailable (Beven et al., 2018b). In such cases, confidence in the forecast relies primarily on the model's ability to reproduce observations within the available range and on the assumption, supported, for example, by physical understanding and expert judgment, that the model remains valid when extrapolated beyond the observed domain. In even more challenging situations, the available observations may be insufficient to perform any formal empirical testing. In these cases, expert judgment plays a fundamental role in assessing whether a given FD or FD-ensemble adequately represents the range of scientifically plausible interpretations relevant to the societal application under consideration (see the lower panel of Fig. 7). Expert elicitation may also be used to assign relative weights to alternative FDs when there is no justification for considering all of them equally plausible (Beven et al., 2018b; Selva et al., 2024). There are also intermediate cases where available data are insufficient for strong tests or were used to build the forecasting model (see box with gradients in Fig. 7); here expert judgment remains essential to interpret and contextualize the results of formal evaluations (Meletti et al., 2021).

We do not purposedly use the term “consensus”, which is widely used in many situations, because it does not have a univocal definition. For instance, in some cases, consensus represents a convergence of evidence or expert opinion toward the same assessment. Conversely, the Italian Civil Protection Code (Italian Republic, 2018) states that the scientific community “participates in the National [Civil Protection] Service by integrating in civil protection activities […] knowledge and products deriving from research and innovation activities, also already available, which have reached a level of maturity and consensus recognised by the scientific community according to the practices in use”. Under this wide perspective, consensus could also be understood as acceptance of procedures and models, not necessarily as agreement on their results (i.e. similarity).

One final point concerns the role of experts' judgment and subjectivity assessment in science (for a detailed discussion see, e.g. Marzocchi and Jordan, 2014; Hanea et al., 2021). The common view that science is inherently objective, whereas subjective judgments fall outside the scientific domain, is misleading. Scientific practice inevitably involves subjective choices and judgments. What ultimately matters is whether probabilistic models and their forecasts (FDs) can be subjected, at least in principle, to rigorous empirical tests of reliability (e.g. validation tests), regardless of the degree of subjectivity involved in their construction.

Paraphrasing Viscusi (1992), communicating uncertainties has never been a popular undertaking. This is particularly true for scientists and decision-makers. Beyond individual facility with public speaking or media, most of them receive little formal training in communication (Besley and Tanner, 2011). To address this, the RETURN project brought in communication experts to share good practices: a total of 32 professionals involved, in various capacities, in risk communication activities were interviewed. Participants operate in different organizational settings, including municipalities, Civil Protection agencies, and public research institutions, or work as freelance consultants. Their perspectives were collected through in-depth qualitative interviews conducted between December 2024–September 2025. The testimonies made it possible to reconstruct established practices, emerging approaches, and ways of conceptualizing and managing complex issues such as uncertainty. Large differences in terminology and skillsets suggest we are still in a pioneering phase of this critical component of effective risk management. No single, universally accepted risk-communication strategy exists, but several practical constraints can help guide efforts.

When discussing communication, “uncertainty” is often overly broad. Here it covers both uncertainty in hazard and risk estimates (probabilities) and uncertainty intrinsic to decision making (Dolce and Di Bucci, 2015): by definition, decision making under uncertainty implies that outcomes cannot always be “right” (van der Bles et al., 2019). Frequently these elements must be combined into a single actionable message (Mileti and Sorensen, 1990). Effective communication therefore must convey the intrinsic scientific aspects of the hazard/risk estimates and the sociopolitical dimensions that shape interpretation, such as cultural values, political interests, decision processes, media, and public discourse (Covello and Sandman, 2001; Slovic, 2000). Consequently, the communication must be not only clear and effective, but also responsible: presented so that it neither misleads nor needlessly alarms, thus enabling informed decisions. To do this, communication should account for variables such as the target audience (with its knowledge level, values, concerns), the decision-making and temporal context, the risks and consequences involved, transparency about knowledge limits, and the choice of channels and language. To achieve this, it must consider all these variables (Sellnow et al., 2017; Sellnow and Sellnow, 2019), as they impact the adoption of protective behaviors, trust in institutions, and community resilience (Paton, 2008).

One major challenge in communicating uncertainty is the generally low level of probabilistic literacy among many stakeholders, including the public, especially concerning the meaning and implications of low-probability, high-impact events. Consequently, communicators must be conscious of their role and capacity: statements made as private individuals, experts, institutional advisers, public officers, or researchers will be interpreted and weighted differently. These roles determine communication objectives, appropriate tools, and accompanying responsibilities. Media channels also differ in how they shape reception, being embedded in everyday cultural routines and practices (Moores, 1993). Thus, tailoring the message to the intended audience is essential. For example, an administration such as the Italian Civil Protection Department must manage both external communication (to the public and press) and internal technical communication within civil-protection structures and the political level. Effective external communication depends on enabling journalists to report accurately, e.g. by providing a dedicated press area during emergencies, releasing data openly, and offering training on scientific uncertainty.

For each responsible institution (e.g. civil protection agencies), communicating with the public entails challenges of expectations and language (Fischhoff, 1995; Warren and Duckett, 2026). Expectations are problematic, because detail is often mistaken for precision: non-experts may demand reliable microscale forecasts and then dismiss as wrong those forecasts that are accurate only at larger scales. Moreover, non-experts (and sometimes scientists) tend to view science deterministically, assuming that researchers will eventually be able to predict every event exactly (Wynne, 1992; Pidgeon and Fischhoff, 2011). Language is also an issue, because science often redefines common words into precise technical meanings that are lost when used externally (for example: extreme or exceptional events, error/mistake, tsunami, moderate alert – and well: uncertainty). Simplifications are sometimes necessary but also risk being misinterpreted: using traffic-light colors to convey risk seems intuitive but can produce over- or underestimation and implies sharp differences for assessments just above or below a threshold. Finally, taking care of social-psychological factors (including bonds with the specific place), as well as preventive ad hoc experiences, are ways to improve uncertainty perception, knowledge and coping by inhabitants (Ariccio et al., 2020, 2021; Stancu et al., 2020, 2025; Villagra, et al., 2021, 2023).

Empirical studies involving professionals of scientific communication identify two approaches to public communication: an “operational” approach based on data, and a “narrative” approach based on storytelling. The former offers strategies to make numbers and data digestible to the interested stakeholder, focusing on the quantitative dimension of science (Morss et al., 2008; Reyna and Brainerd, 2008). The latter uses narrative forms that people naturally use to process information. Narratives can be effective even for communicating very small probabilities associated with important risks: research shows that the understanding of risk improves when numbers are accompanied by familiar, negatively connotated examples or stories (for instance, comparing the probability of an event to the chance that a child has a nut allergy: e.g. Savadori et al., 2022). However, narratives are also risky, because they balance making raw data tangible and comprehensible against conveying subjective or value-laden content that can turn messaging into persuasion or rhetoric (Sellnow and Seeger, 2021).

Communication within a civil protection agency can be aided by the fact that many staff members have a technical-scientific background, but at the same time it is challenging because of their role of link between science and decision-making. Beyond the so-called “strategic” uncertainty that refers to uncertainty about how different actors will behave – for example, scientists may not provide timely scientific information, or politicians may not provide their threshold of acceptable risk (Di Bucci and Savadori, 2018) – the greatest challenge is certainly to define rationale and defensible mitigation actions based on the available scientific information, including its uncertainties. Decision-making often leaves no room for probabilistic nuance, because it operates in a Boolean logic: the output is a definitive decision, yes or no (March, 1994; Renn, 2008). Moving from uncertain scientific results to a definite policy is delicate and requires not only an adequate understanding on the uncertainty, but also sufficient awareness of the competences required, as well as of the responsibilities and legitimacy of the actors involved. In some cases, scientists have clarified their role within the risk-reduction process. For example, Jordan et al. (2014) advocate a hazard–risk separation principle: scientists provide hazard forecasts, sometimes including risk analysis and scenarios impact, while decision-makers set decision-making thresholds, based also on considerations other than hazard and risk modelling, such as feasibility, costs, benefits, and social priorities and potential impact of decisions (Dolce and Di Bucci, 2022). Remarkably, despite its apparent appeal, this principle is often neglected (Marzocchi et al., 2021a).

Tsunami warning systems illustrate this important aspect (Selva et al., 2021). Traditionally, in the NEAM (North-East Atlantic, Mediterranean and connected seas) region, alerts for seismically-induced tsunamis are issued using a decision matrix that assigns alert levels based on the expected characteristics of the triggering earthquake. In other areas of the world, other deterministic approaches are adopted, like best-guess scenarios or envelop models. These systems leave no room for quantifying the inevitable uncertainties, especially given the speed required for alerts in case of small basins, such as the compact Mediterranean Sea and its challenging crustal seismicity. Traditional deterministic methods like decision matrices manage uncertainty by adopting conservative choices, which are related to a certain accuracy and rates of false positives and false negatives; these rates reflect acceptable levels of risk and trade-offs between false alarms and missed alerts, but these trade-offs are often not the result of an explicit choice about the values and political responsibilities at stake. By contrast, a system that uses, for example, a weighted mixture of distributions to provide a probability distribution over possible events can link alert levels to percentiles of that distribution (e.g. Todini, 2017). Choosing percentiles allows calibrating the risk level that triggers an alert and controlling the trade-offs between false positives and false negatives. Representing uncertainty thus makes the value judgments that are the responsibility of decision-makers explicit: when uncertainty is not communicated, choices about acceptable probability thresholds, political by nature, remain implicit.

Another specific feature of uncertainty communication related to risk is whether this communication occurs in ordinary or emergency contexts. Emergencies leave less time for analytical communication and nuanced weighing of uncertainties, yet these are precisely the circumstances when scientific statements may have the greatest impact, i.e. both ignoring and exaggerating uncertainties can have larger consequences (Seeger, 2006; van der Bles et al., 2020). Emergencies also draw greater public attention to science and can often be an opportunity to improve understanding of phenomena and of their uncertainties (Birkland, 1997).

Finally, some emergencies last long enough to allow time for testing the used models; in such cases, ignoring uncertainty can erode public trust in science. In the short term, there is no time to weigh and communicate all uncertainties; in these cases, to avoid the pressure to neglect uncertainties to make messages and actions more comprehensible, the use of protocols facilitates a rapid reaction to the impending peril. These protocols should be defined in advance to leave time for a proper management of substantial uncertainties in the decision-making and communication process. In this way, protocols represent a transparent audit trail, which reduces the degree of subjective opinion in scientific communication with civil authorities and general public. Should a risk mitigation action turn out to be “wrong”, this audit trail would provide a direct means for tracking the decision process in any subsequent formal enquiry. Conveying this message effectively also requires a public education plan.

This paper has summarized the discussion of a multi-disciplinary research group within the RETURN project. The group focused on the importance of disclosing, representing, and communicating all uncertainties in forecasting natural hazards and risks. This perspective favors scientific aspects, not because they matter more than communication, but because the researchers in this group prioritize them as the underlying research is more quantitative and formalised.

Our discussions revealed that different hazard domains face common challenges, regardless of the forecasting models used. Here we summarize our suggestions for overcoming these challenges:

Probabilistic forecasting is the common language for representing uncertainties and bridging science and society. The most impactful events, regardless of their nature, can only be tracked probabilistically over time horizons relevant for rational risk reduction and management measures.

By acknowledging all these aspects, we can achieve our vision of a unified framework for hazard/risk forecasting across all studied phenomena; it will treat hazards across domains in a consistent, probabilistic manner, unifying hazard and risk assessment, model evaluation, and communication.

Calibration of a forecast. Assessment of the comparison between a single representative FD (e.g. a mixture distribution or another combined forecast distribution) and independent observations (see Fig. 6).