Predicting 2016 from June - Satellite Data

Disclaimer

All the temperature sets are in for June 2016, so it's time to update my predictions. But first I want to emphasize that I don't claim to agree with these forecasts. My main intention was to disagree with those claiming back in April that there was 99% certainty that 2016 would set a record. My assumption has always been that the earlier predictions were overstated, and that it would become clear one way or another over the coming months. But at this midway point I think things are even less clear, and may be exaggerating the likelihood of a record even more.

I also want to point out that whilst speculating on which years will be records is interesting, it doesn't tell us much about the trend in global warming. Annual temperatures are fairly arbitrary. A single record does not prove global warming is happening, and a lack of records certainly does not prove warming is not happening.

Roy Spencer of UAH has been talking about the likelihood of UAH beta 6 setting a record as well. In June he was predicting that 2016 will likely be record in the satellite data., but after the June figures showed a large drop he's now suggesting it's unlikely to set a record . So I want to look in more detail at UAH beta 6, which is the least typical of all the data sets. Spoiler - I also think it's unlikely to set a new record this year.

Updates for June

I'm monitoring 7 data sets; the three main terrestrial sets - GISTEMP, NOAA and HadCRUT, and 4 versions of the satellite data - UAH 5.6, UAH beta 6, RSS 3.3, and RSS 4.0. RSS has not so far released version 4 for the lower troposphere so I'm using their product for the total troposphere (TTT), which is the closest to TLT. Here's how the years been progressing for each set (all anomalies based on the period 1981 - 2010).

Most of these data sets showed a drop in June. The exceptions being NOAA and HadCRUT that both showed a small rise.

Here are the graphs showing how the projected year end total has changed throughout the year.

This graph shows the expected annual temperature compared with the previous record year (1998 for satellite data, 2015 for terrestrial data). All of the data sets show a drop in the forecast (compared with the forecast from May), except for HadCRUT which has a small rise. How has this effected the probability of a record year?

Surprisingly there hasn't been much change, with the noticeable exception of UAH beta 6, which drops from 88% to 80%. All others show 90% - 100% certainty. RSS 3.3 and HadCRUT swapped places, with RSS dropping from 94% to 92%, and HadCRUT rising from 92% to 95%. NOAA remains at 98% and the others, GISTEMP and the other satellite versions are greater than 99%.

Summary of Forecasts

Set	Probability	Margin
GISTEMP	1.00	0.16
HadCRUT	0.95	0.10
NOAA	0.98	0.11
RSS 3.3	0.92	0.09
RSS 4.0	1.00	0.20
UAH 5.6	1.00	0.21
UAH beta 6	0.80	0.05

How The Calculations Were Made

In order to get some idea of where these forecasts come from I'll go through the process using UAH beta 6 as an example. For context here's the annual observations along with the forecast (in red) for 2016. The vertical line indicates the 95% interval.

The forecast is to be quite close to the 1998 record, and much warmer than any other year.

The forecast is based on the correlation between the annual temperature and the first 6 months for each year.

This suggests there has been a strong linear correlation between the temperature at the start of the year and the final temperature. There are a few reasons why this correlation shouldn't be a surprise

• The annual figure includes the initial values. In other words, by this point half of the annual figure has already been determined.
• The temperatures for both the start of the year and the whole year are correlated with the underlying trend. That is, because temperatures have increased over time, and years that starts warm are likely to be from the warmer period, and so are more likely to see a warmer end to the year.
• There is some correlation between the start and end of the year even allowing for the underlying trend. That is years that start unusually hot for their time are more likely to remain unusually hot.

The expected temperature can be read straight from that graph. The start of 2016 has been a little warmer than the start of 1998, and 1998 was quite close but a little below the predicted value.

One note of caution is that a fair proportion of the trend line is only supported by two years, 2010 and 1998. On the other hand, it does show that both the strong El Niño years finished up close to their predicted values.

The trend line gives us the expected temperature for the year. This is the temperature we'd expect to get on average, if we had a large number of years all starting with the same temperature. In order to estimate a probability of beating any specific temperature we need to generate a distribution for the range of specific values. This is obtained by calculating the prediction interval (by which I mean I get R to calculate it). I then hack the figures a bit to get the correct scale. The distribution looks like this.

The dotted line shows the 1998 record anomaly, and the shaded are shows where 2016 has to be to beat it. This translates into an 80% probability of beating the 1998 record.

How Plausible are These Forecasts?

Taking another approach (that used by Roy Spencer) we can look at what actually needs to happen in the second half of 2016 to beat the record. For UAH beta 6 the second half of 2016 has to average more than 0.348 °C (using UAH's 1981 - 2010 base period). This compares to the first 6 months of 2016 which was 0.617 °C, and the second half of 1998, which was 0.373 °C.

For 2016 not to be a record would require the biggest difference between the first and second halves of the year in the UAH record. The biggest difference was in 1998 when temperatures dropped by 0.239 °C. By contrast 2016 will need to drop by 0.269 °C - so either way 2016 will set some sort of record!

Here's what 2016's monthly values look like compared with 1998, with the dotted line showing the average temperature needed for the rest of the year to break the record.

Although 2016 started off considerably warmer than 1998, the last two months have shown a much quicker drop than anything seen in 1998. Temperatures will actually need to increase from the June figure to beat the record, but if 2016 reflects the pattern of 1998 temperatures should continue dropping. If this happens there would be no chance of 2016 beating 1998.

It seems likely that temperatures will continue to drop, as they did in 1998, but it's not certain as 2016 is not behaving in the same way as 1998. Overall, based on intuition rather than statistical inference, I suspect the odds on UAH beta 6 setting a record are low - say around 10%.

So why would the statistical analysis be so wrong? In some ways I don't think it is actually wrong, it's just that it predicts the general case. Saying there is an 80% chance of record means that if we had a large number of different years, all starting with the same temperature, 80% would be warmer than 1998, but it doesn't predict which specific years will be in that 80%. You can always bring more knowledge about what a specific year is doing that will suggest if this particular year will be in that 80% or not.

I have tried other statistical approaches, and want to introduce at some point one that gives a better overall performance. This combines the underlying trend with the start of the year, but I'm reluctant to use it at this stage as it actually increases the probability of a record, going in completely the wrong direction to what I expect to happen.

A more accurate statistical method would probably require more advanced methods, such as deep learning, which is outside my experience.

Additional Comments

Whilst it is interesting that despite the strong start to the year UAH beta 6 will probably not set record this year, that tells us little about the state of global warming. For a start we are only talking about one data set, and one that is still unofficial and unpublished. Whilst I'd reduce the odds on all the other data sets breaking a record, they all have more chance than UAH beta 6. My estimates for the other data sets would be that RSS 3.3 has around a 50% chance of setting a record, and the other sets all have 75% or more. (These estimates are as above just my gut feeling.)

UAH beta 6 is dominated by the exceptional value it gives for 1998. 2016 is still likely to be close to the 1998 value and 2016 is very likely to be at least the 2nd warmest year in their history, and is almost certain to be warmer than 2015.

What makes 1998 so unusual is that it came out of much cooler conditions. The two years preceding, and the two years following 1998 were all close to the 1981 - 2010 average. They were warm by 20th century standards, but cold compared to the 21st century. By contrast 2016 follows two very warm years. Possibly one reason this El Niño is dropping more quickly than 1998 is that it has already produced more heat over the previous year. This graph compares monthly values for the three years leading up to 1998 and 2016.

One consequence of this is that the two years 2015 and 2016 are almost certain to be the warmest 2 consecutive years on record, and the three years from 2014 - 2016 will be the warmest 3 consecutive years on record.

The Other Satellite Data Sets

For completeness, here's what the other versions need to do to beat their respective records..

All of them can afford to drop some more and still beat 1998.

Update - 1st August

No sooner had I written the above, when Roy Spencer announces that the UAH beta 6 value for July 2016 showed a slight increase on the June value - up to 0.39 °C.

This hasn't had too much of an impact on my forecasts. The statistical method shows a slight reduction in the probability of this data set breaking a record, down to 78%. I'd still agree with Spencer that UAH will beat the record this year, but the fact that the temperatures went up must increase the odds slightly, if only because it suggests more uncertainty in the temperatures.