As the impact of climate change becomes more apparent in the physical world, the need to accurately assess climate risk is becoming more urgent in the socioeconomic world. When insurers start pulling out of states, as has been happening recently, they frequently assign some blame to climate risk. We rely on climate predictions from models to assess this risk. But how much should we trust model predictions of a distant future?

Weather prediction has well-known scientific constraints that arise from chaos theory. As an initial-value problem, uncertainty in weather forecasts grows over time and eventually overwhelms the predicted signal. We know from direct personal experience not to trust weather forecasts beyond about 10 days, which is close to the 2-week theoretical limit of weather prediction. Bigger computers and more complex weather models cannot take us past this limit.

As a boundary-value problem, climate prediction is fundamentally different from weather prediction. Climate predictions can extend from 10 years to 100 years. But climate prediction uncertainty also grows with time. Are there limits to how far in time and how much detail in space, we can predict? Since we lack personal experience of the validity of long-term climate predictions, we often fail to ask this question when assessing climate risk. This can lead to overconfidence in model capabilities and the belief that with more powerful computers, we can build a “digital twin” of the Earth that can make arbitrarily accurate climate predictions many decades into the future.

Just like weather prediction, climate prediction is also subject to limitations. Some of these limitations are shared with weather prediction, such as nonlinear chaotic uncertainty, and structural model uncertainty, but they impact climate prediction differently. Others, like the uncertainty associated emission scenarios, are specific to climate prediction.

When it comes to quantifying the uncertainties and limits of climate prediction, the devil is in the details: What variable are we trying to predict and over which spatial domain? Predicting temperature change over the whole globe is far less noisy than predicting rainfall in a small region, for example. Coarse-resolution climate models are much better at predicting occurrences of large-scale heat waves than predicting frequencies of small-scale hurricanes.

In this talk, we analyze the limitations of approaches used to quantify the uncertainty of climate prediction. Assessments of risk often pick a single “current-policy” type of scenario as the boundary condition and use a jumble of CMIP climate models and simulations to quantify the uncertainty. The choice of the jumble of models and simulations used for risk assessment is crucial. If not curated carefully, this jumble may not capture the different types of uncertainties properly.