How good have models been at predicting ENSO in the 21st century?

 



In the, roughly, 250 years of the ENSO Blog and our 3.4 million posts, we’ve discussed ad nauseum how complicated El Niño and La Niña are, and how difficult it is to forecast all of the ENSO nuances. Heck, I even wrote a three-part series 75 years ago that evaluated all of our seasonal forecasts (ok, it was 2014-2015 but it still feels that long ago). In a new paper, Azhar Ehsan, friend of the blog and a member of NOAA’s ENSO forecast team, and colleagues analyzed over 20 years’ worth of climate model forecasts of ENSO and found some interesting results.

Why is this paper unique? Well, most seasonal forecasting evaluations focus on model hindcasts, which are forecasts run using past observational data as the start (or initial) condition. For example, if the models are provided the initial conditions on July 1st, 1983, what forecast would it have made? The nice thing about running on past data is that you already know what occurred and can immediately see how well the forecast did. The downside is that sometimes the model development itself can be influenced by this past data. The purest test for models is how well they do in the future, on data that the model has never ever seen. This type of evaluation on “real-time forecasts” is much rarer, and is exactly what Azhar and his co-authors did.

ENSO Terms and conditions

La Niña and El Niño make up the El Niño/Southern Oscillation, or ENSO. La Niña is characterized by cooler-than-average sea surface temperatures across the central and eastern tropical Pacific Ocean. El Niño is the opposite phase, with warmer-than-average water present across the tropical Pacific Ocean. These changes in sea surface temperature across the Pacific jumble up the atmosphere above which can lead to global impacts on climate patterns. Seems pretty important, right? And unlike most other climate phenomena, the state of ENSO can be forecast months in advance, giving communities time to prepare.

Tell me about that sweet, sweet data

Let me paint you a picture. It’s February 2002. Crossroads starring Britney Spears has just come out, while Ja Rule and Nickelback are burning up the music charts. At Columbia University’s International Research Institute for Climate and Society (IRI), a plucky group of scientists has begun issuing ENSO forecasts. That effort has now become the world’s longest archive of real-time monthly ENSO forecasts from modeling groups across the globe. The list of forecast contributors has continued to grow since 2002, and the tally of the treasure trove of climate model data currently stands at 28 different climate models.

Why so many? No single model forecast is ever going to be exactly correct all the time. To get a sense of the range of potential outcomes, it’s important to not only have a bunch of forecasts using the same climate model due to the chaos of climate but also forecasts from a bunch of different models due to the idiosyncrasies of each individual (we call these combined bunches multi-model ensembles). A well-constructed forecast ensemble won’t tell you precisely what outcome to expect, but it will tell you how much the odds are tilted toward one outcome or the other (i.e., probabilities).

The model forecasts can be split into two types, dynamical and statistical. Dynamical models refer to models which take observational data to simulate earth’s future climate by using equations that represent our best understanding of the laws of physics (e.g., like the computer climate models that make up the North American Multi-Model Ensemble, or NMME, that are frequently featured in this blog). Statistical models, on the other hand, use the historical relationships between ENSO and other climate variables from the observational record and then use these relationships to make predictions for the current situation.

 

📌 Visit Us 👇

Comments

Popular posts from this blog

Data experts race to preserve US government statistics amid quiet purges

11 Essential Statistical Tools for Data-Driven Research

Trump Gets Rid of Those Pesky Statistics