What a record July means for 2023 temperature

Just how likely is 2023 to be the warmest year on record?

Last month was the warmest on record by a large margin in reanalysis products like ERA5, JRA-55, and NCEP CFSv2. We expect to see a record by a similar margin in the global surface temperature datasets from NASA, NOAA, Berkeley Earth, and Hadley that will be released in the coming weeks.

A few weeks ago, I produced an analysis for Carbon Brief suggesting that, based on the year-to-date and El Nino forecast, 2023 was now more likely than not to be the warmest year on record. I noted at the time that the odds presented were likely overly conservative, as they “did not include the extremely warm July temperatures seen in daily reanalysis products.”

Now that July is in (and set a new record by an exceptional margin), we can take a look at the odds of 2023 being the warmest year on record, and whether there is a chance that 2023 might pass 1.5C above preindustrial for the year as a whole (which, its worth mentioning, is not the same as what most people mean when they talk about “passing 1.5C”, which the IPCC defines as exceeding the 20-year average).

Let’s start by comparing 2023 to-date in the ERA5 dataset to the current record holder, 2016. The figure below shows 2023 in red, 2016 in blue, and all other years as grey lines.

Monthly ERA5 temperature anomalies between 1940 and 2023.

If we examine average temperatures for the year-to-date, 2023 is in third place at 0.62C relative to the 1981-2010 baseline period, behind 2020 (0.67C) and 2016 (0.68C).

However, 2023 had something of a cool start, and with a strong El Nino event building in the tropical Pacific we expect temperatures to remain elevated for the remainder of the year. This increases the odds that 2023 will beat the prior 2016 record compared to a year where we had ENSO (El Nino and La Nina) neutral conditions.

Calculating the odds

So how do we determine the odds that 2023 will be the warmest year on record? We can turn to the old workhorse of science: multiple linear regression.

Specifically, we can use the relationship between the temperatures of the prior year, the year-to-date, and the forecasted ENSO conditions to predict the central estimate of 2023 annual temperatures along with the two-sigma uncertainty. This results in the regression equation below:

era5_annual ~ last_year + year_to_date + first_7_enso + last_5_enso_forecast if year != 1982 & year != 1992

The resulting estimates for 2023 are plotted in the figure below, along with their two-sigma (~95%) uncertainty range. This figure shows global annual temperatures from ERA5 relative to a preindustrial (1850-1899) baseline, using the HadCRUT5 dataset to fill in the early part of the record (and aligning anomalies between the two over the 1981-2010 period).

Predicted 2023 annual temperature in ERA5 based on year-to-date and forecasted ENSO state.

Using this approach, we find a best-estimate of 1.38C, with a 5th to 95th percentile uncertainty range of 1.30C to 1.46C. This gives us an 85% chance of 2023 being the warmest year on record. Note that the uncertainties might be a bit smaller than they otherwise would, as the model assumes perfect knowledge of future ENSO states (while in reality there is a pretty wide envelope of ENSO projections across models).

We can also use this linear regression to see how the actual values for each prior year compare to the regression predictions, using the same variables (year-to-date, prior year, first 7 month ENSO conditions, last 5 month ENSO forecast).

Predictions of prior year temperatures based on the prior year, first 7 months of the year, and ENSO state.

Of course, the choice of parameters to use in linear regressions can be more of an art than a science at times. There is not necessarily a clear, objective set of predictors, and the choice of which to include can affect the results.

For example, one could argue that the most recent month’s temperature anomaly provides useful information for the model. After all, there is some autocorrelation in temperature anomalies, and July may tell us more about what to expect for the remainder of the year than January. Lets try a variant of the model that also uses the latest month as a predictor:

era5_annual ~ last_year + year_to_date + latest_month + first_7_enso + last_5_enso_forecast if year != 1982 & year != 1992

Here we see that the inclusion of the latest month as a predictor significantly increases the predicted 2023 annual temperature.

Predicted 2023 annual temperature in ERA5 based on year-to-date, the July anomaly, and forecasted ENSO state.

This model gives a 98% chance of 2023 being the warmest year on record. The central estimate is 1.42C above 1850-1899 levels, with a range of 1.35C to 1.50C. It suggests there is a ~2.5% chance that 2023 exceeds 1.5C, and turns out to have a slightly better fit to the data (adjusted r²) than the version that doesn’t include the latest month. It also produced a good hindcast.

Predictions of prior year temperatures based on the prior year, first 7 months of the year, July anomaly, and ENSO state.

So are the extremes we’ve seen in June and July indicative of the remainder of the year to come? Or are will we return to mere record warm temperatures for the remainder of the year? There is no clear right answer here; I’d suggest they are likely to be more representative of the remainder of the year than the early months, but also hold out the possibility that we’re seeing some anomalous internal variability that might dissipate.

Either way, at this point the odds are heavily in favor of 2023 being the warmest year on record. And given the continuing development of strong El Nino conditions in the tropical Pacific, its likely that 2024 will be even warmer.