Code and data for this post can be downloaded here. (It’s getting sloppy, I know…I’m not an R guy).
In my last post, I looked at the 20th century hindcast within the context of assuming the multi-model mean represents the true forced component of our climate, and the weather noise/errors can be simulated using an AR(1) process. With the two charts below, I’ve extended this to look at the first 11 years in the 21st century, and also included the GISSTemp and NOAA calculated anomalies in addition to HadCRUT.
As you can see, the differences between the individual observational series are tiny compared to the differences between our various “runs”. The following zooms in closer to focus on the period since 1980, although it maintains the 1900-1950 baseline:
One thing that stands out is that recent years have fallen increasingly below the multi-model mean in all 3 observational series, even landing outside the 54 pseudo-runs in places. On the one hand, this seems unique to the 21st century, suggesting the multi-model mean isn’t doing as good of a job of forecasting as it has in hindcasting. On the other hand, the degree to which current observations fall below the multi-model mean is sensitive to the baseline chosen – I’ve used 1900-1950 for similarity to the IPCC AR4 report, but a more recent baseline will lessen the difference. So it’s worth noting that a slight overestimate of warming by the MMM in the latter part of the 20th century hindcast has also contributed to this difference. This will become more evident when looking at the histograms of trends.
So, what if we compare the trends of the observations to our MMM + AR(1) noise model? Each histogram below is of 1000 run of the MMM + our AR(1) process, using the same parameters as above. The first chart is a similar to one that can be found at Lucia’s here, except that it uses the 1000 “pseudo-runs” instead of only the 54 runs available in the ensemble, and it uses annual averages that must end in 2010 rather than going up to the current month.
Based on this model, there is quite a bit of variation among the individual runs, since there are only 10 years/points. It’s also worth noting that 2010 was an El Nino year, and that the relatively cooler months of 2011 will not be included in the trend calculations above until the end of the year, which could eventually result in lower percentiles for the observational series (especially given another point in the MMM predicting warming).
Extending back to 1980, we see some of the weather noise appears to disappear from the trends (as expected), thereby tightening the confidence intervals:
Finally, I mentioned before that the MMM overestimation compared to observed warming seems to have been present even in the hindcast, particularly in the latter part of the century. The graph below shows trends from 1960 with this MMM + AR(1) model.