Written by Yazmina Zurita, physicist and R&D Data Scientist at Nebbo.
Predicting weather involves some uncertainty, as the Earth system is inherently chaotic. This means that even tiny differences in initial conditions can lead to vastly different outcomes as time progresses. This unpredictability is particularly significant in seasonal forecasts that aim to predict weather conditions months into the future. Variations in starting conditions, such as those caused by rounding in calculations, contribute to uncertainties. Moreover, no model can perfectly emulate the complex realities of the natural world because of necessary simplifications, computational limits, incomplete knowledge, and measurement inaccuracies. These elements collectively enlarge the uncertainty in predictions.
Although it is challenging to quantify the uncertainty each factor contributes to a forecast precisely, we can approximate the total uncertainty by utilizing confidence intervals.
Confidence intervals and their Calculation
A confidence interval provides a range of values where the actual result is likely to fall. Along with the interval’s boundaries, there’s a component known as the confidence level. This is a percentage indicating how often the result is expected to fall within the specified range. For example, if a future wind speed is estimated to be between 2 and 5 m/s with a 90% confidence level, it implies that under similar conditions, the observed wind speed will fall within that range 90% of the time.
Various methods exist for quantifying uncertainty. One such method is conformal prediction, a machine learning approach that generates statistically valid and adaptable confidence intervals by analyzing past errors between predicted and observed outcomes.
Assessing the Quality of Confidence Intervals
Two key metrics assess the quality of confidence intervals: coverage and width. Coverage refers to how often the observed value actually falls within the interval, while width refers to the breadth of the interval itself.
The ideal scenario is for coverage to match the confidence level. Regarding the interval width, narrower intervals are preferable as they represent a smaller range of potential outcomes, indicating greater prediction precision.
There is a trade-off between achieving high coverage and maintaining a narrow interval. A higher coverage level ensures more certainty about potential outcomes, but it requires a broader interval to account for more possibilities. Thus, a higher confidence level results in a wider interval.
The Value of Confidence Intervals in Forecasting
In conclusion, confidence intervals are invaluable for estimating uncertainty. Reliable methods exist to construct them, and metrics are available to evaluate them. Moreover, the confidence level can be adjusted according to individual risk tolerance. Although these intervals can be wide, particularly for long-term seasonal forecasts, there are periods when they become narrower. In these instances, the prediction intervals provide a precise range of likely outcomes with a customizable degree of certainty.
