Extreme Event

Most things in life move within a familiar range. The stock price fluctuates by a few percent, the river rises a bit in spring, electricity consumption is higher in the evening than at night. An extreme event is an occurrence that lies far outside this familiar range. It happens rarely, often only once every few decades, but causes enormous damage. Examples include a thirty percent stock price crash in a single day, a flood the likes of which hasn’t occurred in a hundred years, or a widespread power outage. The term is deliberately broad and is used equally in finance, climate research, and engineering.

Why rare cases cause the greatest damage

Those who assess risk usually look at the average. That is precisely the wrong perspective when it comes to extreme events. The average stock market day reveals nothing about the one day on which a fund loses half its value. A single such event can destroy more money than ten good years brought in.

On top of that comes a very common fallacy. Because something has never happened, many people consider it impossible. In fact, it only means the observation period was too short. The statistician Nassim Taleb coined the image of the black swan for this: in Europe, all swans were thought to be white, until black ones turned up in Australia.

For banks and insurers, this is not a thought experiment but a regulatory requirement. Supervisory authorities demand that institutions also calculate for severe crisis scenarios. These calculations are called stress tests. They do not ask what will probably happen, but what would happen in the worst case.

How to make the rare calculable

The core problem is the lack of examples. An AI model learns from past data. If a case appears only once or not at all in that data, the model cannot reliably recognize it. Instead, it optimizes for the normal case, since that is where most of the points can be gained. A fraud detector that, given a fraud rate of 0.1 percent, simply always says “no fraud” is correct in 99.9 percent of cases and yet is useless.

Statistics has its own tool for this: extreme value theory. It does not look at all the data, but only at the highest deviations, such as the annual peak levels of a river. Separate mathematical distributions apply to these deviations. This makes it possible to estimate how high a water level might rise once every hundred years, even though only fifty years of measurement data are available.

In AI, two additional tricks help. Artificial examples of rare cases are generated so the model has something to learn from in the first place. And errors are weighted differently: a missed extreme event costs far more penalty points during training than a false alarm. Both shift the model’s attention from the average toward the edge.

From flood protection to the data center

In financial news, the term usually appears as tail risk. This refers to the outer edge of the probability curve, where the rare, costly cases sit. Metrics such as Value at Risk attempt to capture this risk in a single number. After the 2008 financial crisis, this was seen as a serious misjudgment: the models considered the collapse to be practically impossible.

In climate research, the focus is on heat waves, heavy rainfall, and droughts. Here the situation has shifted, because warming makes some extremes significantly more frequent. Events that used to occur once every fifty years now occur once every ten years in some places. Historical data alone is therefore no longer sufficient for planning dikes or sewer networks.

The topic is also present in engineering. A self-driving car must be able to handle a situation it has never seen during training. Data center operators plan for the simultaneous failure of multiple systems. In both cases, the same rule applies: it is not normal operation that determines safety, but the rare bad day.