After major disasters, public debate often treats them as unexpected or unprecedented. This reaction is not necessarily about the absence of warnings. It reflects how societies process shock – and how authorities often explain disruption as unavoidable, rather than the result of earlier choices.
Extreme weather is rarely unpredictable. Days, sometimes weeks, in advance, scientists are able to warn of an increased risk of storms, floods, droughts or other hazards. Yet the cycle repeats.
To understand why this is, colleagues and I reconstructed the scientific warnings and the official responses to major floods in Luxembourg in July 2021 – my home country’s most damaging disaster on record. Those floods caused far more damage than they would have done if early action was taken, but Luxembourg isn’t an outlier: many other countries suffer from the same problems we identify.
As the UN targets “early warning for all” by 2027, it’s worth noting the issue is not that warnings were missing. It is that warning systems are often designed to act on certainty rather than probability – and that’s not how forecasting works. By the time warnings become visible to the public, it is often too late.
Abaca Press / Alamy
Weather forecasts may look definitive on your phone, but they are probabilistic by nature. They are created by running a series of computer simulations of the future weather. The level to which the outcomes of different simulations agree with each other provides the likelihood of hazardous conditions, not guaranteed outcomes. These allow forecasters to identify elevated risk well before impacts occur, even if the precise location of an event and their size remain uncertain.
Crucially, uncertainty is usually greatest further ahead, when preventative action would be most effective. Acting early therefore almost always means acting without certainty. This is not a weakness of science, but an inherent feature of anticipating complex systems under changing conditions. The real challenge lies in how institutions are organised to interpret, trust and act on those probabilities.
Acting on certainty
Most warning systems rely on predefined procedural thresholds: alert levels, activation protocols and emergency plans that kick in once specific criteria are met. Forecasting may indicate that flooding is increasingly likely, for example, but measures such as evacuations or road closures can only be triggered after formal thresholds are crossed.
Before that point, risk information passes through many layers of interpretation and judgment, where early signals are often noted but not acted upon.
C3S/ECMWF (Data: ERA5), CC BY-SA
Thresholds serve important purposes. They help coordinate response, clarify chains of command and reduce unnecessary disruption. But they also embed a structural preference for certainty. Action is authorised only once risk is framed as imminent, even when credible evidence already points to escalating danger.
This attitude was apparent in the days leading up to the July 2021 floods. Our study shows that multiple forecasts at European and national levels indicated a high probability of extreme rainfall and flooding, in some cases up to a week in advance. This information was available across different parts of the warning system. At that stage, uncertainty about precise impacts remained, as would be expected. What mattered was how the system was designed to handle that uncertainty.
Too early for warning
Because Luxembourg’s response measures were tied to procedural thresholds, early signals could not translate into anticipatory action. The country’s water administration and its national weather service had access to relevant information, but they operated within a framework that did not authorise a collective interpretation of what was happening or encourage action before thresholds were crossed.
This was not a scientific miscalculation, nor was it necessarily an operational mistake by individual agencies. Meteorological and hydrological services most likely did as much as their mandates allowed. The decision to wait for formal triggers was human and institutional rather than technical,
reflecting a system designed to prioritise procedural certainty over sound decision-making.
Copernicus EMS/ECMWF, CC BY-SA
By the time action was authorised, for many people it was too late. Evacuations or installing flood gates became far more difficult, particularly for communities with limited experience of such severe floods. From the perspective of those affected, warnings appeared late or did not arrive at all – even though the risks had been identified earlier throughout the system.
Luxembourg is a particularly instructive illustration of what can go wrong, because it is a small, wealthy and well-connected country. The issue was not necessarily a lack of resources or scientific capacity, but of institutional design and societal readiness to act on risk.
Learning and resilience
The effectiveness of early warning systems over time depends on their ability to learn from extreme events. This requires open, independent analysis of what worked, what did not work and why. In several neighbouring countries affected in 2021, such as Germany and Belgium, formal inquiries and external reviews were carried out. In Luxembourg, they were not.
When expert critique is discouraged or avoided, learning slows. Questions about system performance remain unresolved and the same structural vulnerabilities are likely to persist. This creates a systemic risk in its own right: societies become less able to adapt warning systems, interpret uncertainty and act earlier on emerging threats.
As someone who has worked within these systems and continues to research disaster risk governance, I have seen how asking difficult questions can be treated as destabilising rather than constructive. Resilience depends on confronting uncomfortable truths, not avoiding them.
The risk of extreme weather is increasing across Europe and beyond. Early warning systems are rightly central to disaster risk reduction. But their effectiveness depends on how societies authorise action under uncertainty. This is a choice, not an inevitability.
Uncertainty cannot be eliminated. The challenge is to decide how much uncertainty is acceptable when lives and livelihoods are at stake. Systems designed to wait for certainty – for procedural, organisational, financial or reputational reasons – are more likely to deliver warnings that arrive too late to feel like warnings at all.
If resilience to future climate risks is to be sustainable, warning systems must be designed to learn, adapt and act earlier on credible risk.
