After I posted GISS Spackle and Caulk, a number of commenters marveled at the symmetry of the histogram (GISS temperature estimate minus actual temperature). Some were dismayed that there was not a clear warming bias in the plot. Others were giddy for the very same reason. A few noted (as I hoped) that the differences tended to be rather large, but most seemed content with the fact GISS could hit the side of a barn from five feet.
No one should be surprised with the shape of the histogram. The “simulation” I performed required that all three months be available in a specific season for a specific station in order to calculate an estimate and compare it to the real value. For example, if summer 1957 was being tested, I needed June, July and August. If August were missing, the GISS algorithm would not be able to estimate June or July, and I would not have a real August to look at either.
With all three months available, I forced symmetry into the result. For every over-estimated August I needed a corresponding under-estimated June or July. The algorithm demanded that if I estimate all three months, their average must match the true average.
However, that is akin to saying that if I flip a coin often enough, the number of heads will be roughly equal to the number of tails. As most of us have experienced, coin flipping can be quite streaky. It is not uncommon to flip eight heads in a row. But having flipped that many heads does not change the probability of the next coin flip.
And so it goes with temperatures. In the actual application of the GISS algorithm, at most one month in a season can be estimated, so symmetry is not guaranteed. If one month is estimated more than another, it might be possible to introduce asymmetry.
As chance would have it, one specific month-year GHCN entry has had its temperature estimated by GISS far more than any other combination in the record. And as luck would have it, we have real GHCN data to compare against those estimates.
As has been noted repeatedly on this blog, MCDW records began replacing most non-US temperature records in the late 1980s. Most of the MCDW records begin in January 1987, and the records they replace generally end between December 1989 and December 1990. During the period of overlap, the MCDW records usually match exactly those that they replace. In a few cases they might differ by one or two tenths of a degree in the occasional month.
When an MCDW month begins in January 1987, the winter season temperature (DJF) is missing the December 1986 value, so GISS must estimate it. But the record that MCDW replaces contains a real, live December value. This means that, when an MCDW record agrees with an existing record during the period of overlap, the real December value can be compared against the estimate from MCDW. So of course I looked at this for all GHCN records.
Following is a histogram showing the GISS estimate of December 1986 minus the actual for GHCN stations in Europe and Russia. I will show other regions of the world in future posts. The reason I focus on this broad swath of land is that the resulting records are among the most lengthy available.
One might notice GISS under-estimates December 1986 for this region by a greater than 2 to 1 margin. So how does that affect the temperature record? (The resolution of the histogram is 0.5 degrees. There are actually only three exact matches to the temperature record. The remaining estimates in the “0″ bin are actually colder by a 2:1 margin).
When GISS combines multiple records for a single station, it uses an undocumented enhancement to the “bias method”. This “enhancement” starts with the latest record, that being the MCDW record, rather than the documented longest record, which is usually the one with the real December 1986 temperature. An average temperature is calculated for both the MCDW record and the older record for the period of overlap, which is usually 1987 to 1990.
If the average temperature of the older record during the period of overlap is warmer than the MCDW record, the older record is “cooled” to match MCDW. The opposite is true if the older record is warmed to match the MCDW.
All things being equal, a cold estimate for December 1986 results in a colder 1987 to 1990 MCDW, as compared to the older record. Therefore, the older record is uniformly cooled to match the MCDW estimate. June 1906 is cooled just as much as February 1985. Across Europe, Greenland, Iceland, and all of Russia this happens more than twice as much as the warming does. And it is done to many records that go well back in time.
By cooling the older record and leaving the current record unchanged, an enhanced warming trend is introduced. This is completely artificial, of course, because the actual December 1986 temperatures are available.
In this case, GISS got their eight heads in a row.