Predicting 2016 Temperatures - Part 1

Introduction

Most of the global temperature sets have been published up to April 2015 and they all continue to show the globe is very warm. GISS in particular has had 7 months of record breaking anomalies, with every month since October 2015 being more than a degree above the base line average (1951 - 1980).

Of course, this spike in temperatures is generated by the current El Niño conditions, and temperatures will fall back later in the year.

This leaves a question as to whether 2016 will be another record breaking year, in some or all data sets. Personally I don't care too much for emphasizing annual records, as it's a distraction from the long term trends - a single record warm year does not prove the trend is upwards, and a lack of records over a certain period does not mean warming has stopped. Nevertheless, records are fun and if 2016 is a record it will be remarkable given that it will be the 3rd record year in a row for land based observations. For the satellites the question is whether 2016 will finally break the long standing record set in 1998.

It's something of a risky move predicting a record, given how the doubters will use anything you say in evidence against you - If you predict a record that doesn't happen it will be prove that all your forecasts are wrong, in it does happen it will be evidence that the figures are fraudulent. So I was intrigued by a couple of tweets Gavin Schmidt made over the last couple of months, saying that it was more than 99% certain that 2016 would be a record at least as far as GISS goes.

Aside from the wisdom of making himself such a hostage to fortune, I was also suspicious of the idea that 3 or 4 months in you could be so confident. So I've been looking at the statistics myself. This proved to be so interesting, that I think I might try to do a month by month summary.

Simple Correlation

The first question is how much correlation has there been in the past between the first 4 months of the year and the final Annual temperature.

The correlation is remarkably good, and looks pretty linear. I was surprised at how strong the correlation was, but on reflection there are a couple of reasons why this shouldn't be so surprising. First, by April a third of the year has already been locked-in. More importantly as temperatures have warmed in general it's natural that this will increase both the annual average and the starting average.

Using this line to predict 2016, we get a forecast of 1.08 ° C, with a 95% prediction interval of 0.94 - 1.23. This compares with the current record set all the way back in 2015 of 0.87 ° C. From this I estimate the probability of 2016 setting a record in GISTEMP at 99.8%.

This graph puts the forecast in the context of previous annual temperatures. The vertical line represents the 95% prediction interval.

Is this a reasonable prediction? I'd be extremely cautious about reading to much into this simple analysis. For one thing, extrapolating from a trend is dangerous when you move outside the range from which the trend was calculated. In this case the average for Jan-Apr is far warmer than anything seen before, so it's impossible to know if the trend would continue linearly. The red dot in the next graph shows where our prediction sits on the line.

In addition any probability is only the probability of a record assuming the assumptions of the model are correct. There will always be a lot of factors that such a simple model does not take into account, and given the probability is so high, there's a reasonable chance that any additional factors will reduce that probability. One thing in particular is that this model does not take into account the fact that this is an El Niño year, and that it is almost certain that temperatures will drop during the rest of the year. Though looking at the last big El Niño year, 1998, that ended up very close to the predicted value. I want to look at more complicated forecast models in a later post.

All of the above has been with regard to the GISS temperature set, but we can use the same method to look at other sets.

For NOAA the probability of a record is slightly less, partly because 2015 was somewhat warmer, 0.9 ° C, and partly because the start to 2016 hasn't been quite as warm in NOAA as in GISS. The prediction for NOAA using this method is 1.01 ° C, with a probability of 97.9% of beating the 2015 record.

For HadCRUT4 the probability of a record year is only 91.7%.

For the satellite data the current record goes back to 1998, and the two versions of data sets used which show the smallest amount of warming, RSS 3.3 and UAH beta 6, show the most uncertainty. RSS 3.3 is projected to beat 1998 by 0.13 ° C, with a 95.5% of a record. UAH beta 6 is projected to beat 1998 by 0.08 ° C, with a 87.0% chance of a record.

By contrast UAH v5.6 has a 99.9% chance of being a record, and RSS TTT 4.0 has a 99.7% chance. It's curious that for UAH the newer, but still unpublished, version is the data set with the greatest chance of not being a record, but the older official version has the most chance of any set of beating the record.

Here's what the UAH beta 6 forecast looks like in context.

This table summarizes the probabilities and margins for all data sets. I'm showing the expected value as a margin over the previous record, rather than as an anomaly to avoid confusion between the different bases used for each set.

Set	Probability	Margin
GISTEMP	0.998	0.22
HadCRUT4	0.917	0.12
NOAA	0.979	0.15
RSS 3.3	0.955	0.13
RSS 4.0 (TTT)	0.997	0.25
UAH 5.6	0.999	0.26
UAH Beta 6	0.870	0.08

Changes over time

Here's a graph that shows how the probabilities have changed since the January figures.

and here's a graph showing projected difference between 2016 and the previous record in degrees C.

Conclusion

I think it's very likely that most and probably all data sets will show 2016 as being the warmest year on record. But I would be pretty skeptical about the very high probabilities obtained by this simple method. All of the above statistics should be considered just for fun and I take no responsibility for any losses occurred by anyone taking bets. My main interest in all this is to see how the forecasts, using this simple method, change over the year.

Update

This post was updated on June 1st to include HadCRUT4 figures for April.