Skip to content

Greenland’s Climatic Shift In 1922

October 22, 2016

By Paul Homewood




I took a look at Bo Vinther’s SW Greenland temperature series the other day. Robin Edwards, who alerted me to it, has also written an analysis of it, and his guest post is below.

It gets a bit technical, but is worth reading through to the end:





Guest Post by Robin Edwards



"Consider an analytical technique that imposes no preconditions on the data, one that accepts all viable observational work as being honest and painstaking, is little affected by randomly occurring missing data, or even the occasional "bad value" which might be due to transcription problems.  The technique could provide information on possible changes of level and the occurrence of very rapid, short-lived, change which would be obscured by the usual technique of naive linear fitting, or disguised by smoothing procedures of any sort.


In the 1950s a statistical quality control technique was introduced with the major aim of responding rapidly to a change in an output parameter of a production process.  It is known as the Cumulative Sum technique (Cusum), and had the important property of being very simple to calculate without computing equipment – essential at the time of its development.  This technique can of course be applied to historical data to explore retrospectively the behaviour of time series.


What is a cusum?  Very simply, the cusum of a series is formed by subtracting a suitable constant from every observation, and successively summing these differences – thus the name cumulative sum.  It has also been called "cumulative deviation from the mean" by some workers in particular contexts.  Generally the most suitable constant for retrospective analysis is indeed is the mean of the series under scrutiny, which ensures that the plotting scale will be a comfortable one, with the cusum ending at zero. The "technology" has been used in industrial quality control for 50 or more years as a means of spotting changes in level of a product quality parameter.  It is useful in this role, but there are severe technical objections to using it as a forecasting tool, due to its obvious serial correlation.  This is not worrisome with climate data since the method itself shows that forecasting (projecting) climate from its past behaviour is likely to be impossible. Rather, it seems to support the view that weather and climate are inherently chaotic. This does not affect its use as a tool as method for examining historical data, for which such an objection is irrelevant.


The cusum series is plotted against observation number or more usually time.  This plot has some very interesting and simple properties.  First, a roughly constant sequence in the original data results in a (roughly) straight segment of cusum. Furthermore, its angle to the x axis is a representation of the typical difference between the reference value and the value prevailing over the straight segment.  Any abrupt change in the slope of a cusum indicates a rapid or perhaps step change in the original data.  This is a vital fact that sets apart the cusum from other data plotting methods.  The slope of the cusum plot is the prevailing difference between the basis value (usually the average of the segment under examination). To repeat, by adding the geometric cusum slope to the base value you arrive at the temperature prevailing during the period being studied.  Very importantly, a visually linear or smoothly curved cusum identifies a data segment that merits examination by linear regression (fitting the least squares straight line) for the original data to establish its important parameters, normally its mean, standard deviation and slope. With truly linear cusum we would expect the resulting linear coefficient to be indistinguishable from zero. In such cases the inferential statistics will confirm that there is no significant slope over that period, but this does not always happen.  This is because the choice of range – made by visual inspection of the cusum plot – may cover a period of real change that was not readily apparent in the cusum plot.  Smoothly curved segments of cusum plots are a sure indicator of steady change in the original data, and can be used to identify regions of the original data that are well suited to linear fitting, with the computing as yet unquantified trend as its objective.  In such cases the t statistic for the slope of the regression will indicate whether a significant slope (at some chosen probability level) exists.


Thus, this approach uses the cusum plot to identify stable regions which may subsequently legitimately be examined by the technique of linear fitting.  These regions may be constant or have a quantifiable linear trend – and using the cusum to identify them avoids the common pitfall of carrying out a trend analysis over a region where the underlying data are far from linear in character.


A very useful feature of this form of cusum analysis is that it appears to be little affected by missing data, even fairly substantial omissions.  Because the primary purpose of cusum analysis is to identify major enduring sequences of data so that they can be further examined by conventional fitting methods, "poor" or missing data that change the appearance of cumulative sum only marginally do not in general influence the identification of interesting time segments."


Cumulative Sums have by their nature a huge autocorrelation and cannot be applied in any sort of projection of climate other than a very general one.  They are not based on an hypothesised model, which would be required for any sort of "forecast" to be made. Judging from past experience, any attempt to make climate projections over a long period is doomed to abrupt failure.  The word "abrupt" is used deliberately, since it describes a common feature of climate as analysed by cusum technology.  This is that sudden changes seem to occur randomly and with no preliminary indication that they are about to happen.  This is in marked contrast to most analyses, which seem to be intended to downplay or eliminate any chance of recognising the presence of a sudden change.  One has to assume that such a notion lies outside the range of possibilities that climatologists are prepared to contemplate.  "Forcings" such as greenhouse gas concentrations or orbital effects change very gradually, and climate models that incorporate them as essential elements of the modelling procedure are likewise constrained to move in a gradual (imperceptible?) fashion.  Conventional analyses are well suited to this state of affairs, so it is possible that little effort has been made to identify sudden changes.  Volcanos however do provide a sudden and short-lived forcing, and apparently can be recognised in hindcasting by climate models, although I have no reference for this.  Cusum technology usually identifies them fairly readily, but of course cannot forecast them.  However, unsuspected sudden change seems invariably to be missed by the standard climate models even retrospectively, whilst the cusum approach makes them obvious.


Thus, unsuspected abrupt change is very unlikely to form a part of a model’s output. Essential input parameters are unlikely to be available.  Its name tells one why.  By its nature it cannot be foreseen.  However, this does not mean that it does not happen. Working with cusums provides the opportunity to demonstrate retrospectively the possibility of the reality of step changes.  I believe that climate models cannot by their nature identify them, and thus ignore them.


So, here’s a very small selection of GIFs which might help to demonstrate the cusum approach in action. I’ll attach SWGrn(land) monthly averages taken from Vinther’s paper. Then the transform to monthly differences, which eliminates the gross seasonal factor. You will see that the range of the data is much reduced (of course) and that a structure has appeared.  Now I’ll attach the cusum of the Annual Averages of the Vinther data.  You will see the gross discontinuity even in this diagram, but its time of occurrence is made more precise by using the cusum of monthly differences. You can see that I made this diagram in 2008. Vinther was not interested!


Without going further at this time it is clear to me that some simple inferences can be drawn from this MDCusum. First, that the period up to 1922 was quite stable, with a gradual increase of temperature (the positive curvature). Second, some drastic happened in 1922 (September, if one makes a detailed plot), in that the cusum has a very abrupt change in slope. As I’ve indicated above this is simply a step change in the data.  The difference in the slopes of the pre and post 1922 cusum regimes is the numerical size of the step. Third, up to approximately the 1970s the cusum is remarkable stable, showing that the original data were effectively constant. Fourth, since then the cusum is very variable, indicating unstable conditions – which can easily be examined by more local cusum calculations.










All data is from Vinter’s paper, “Extending Greenland temperature records into the late eighteenth century”




PH Thoughts


One of the enduring myths of climate science is that temperatures in Greenland have been rising alarmingly recently.

As Robin’s cusums show very clearly, the real climatic shift occurred in 1922.

Real scientists have known about this for many years. Kenneth Drinkwater, for instance, wrote a paper in 2006, “The Regime Shift  of the 1920s and 1930s in the North Atlantic”, which stated during the 1920s and 1930s, there was a dramatic warming of the northern North Atlantic Ocean.

This phenomenon was not limited to SW Greenland, but was seen across a wide swathe of the Arctic from Canada in the west to Siberia in the east.

It was also in 1922 that NOAA reported a radical change in climatic conditions, and hitherto unheard of high temperatures in Spitzbergen.

But telling the truth does not get much grant funding nowadays.

  1. R2Dtoo permalink
    October 22, 2016 3:48 pm

    A very interesting analysis. If this can be done for many arctic locations, and the pattern is consistent (as suggested), it would be useful to put it out and have the science network speculate on why it happened. Ostensibly, this step-change predates human influence vis-a-vis greenhouse gases. IIRC another step-change occurs in the late (?) 70s. I believe Bob Tisdale shows similar patterns. If global, or even regional temps, move in steps, linear analysis will hide the pattern, and will be as useless as it is if patterns (long-term) follow a sine wave.

  2. Broadlands permalink
    October 22, 2016 7:42 pm

    Temperature and barometer measurements were made monthly at Nain, Labrador in eastern Canada in 1779-1780. The yearly mean was 30.3°F. The current “normal” at Nain is 26.65°F. An ‘alarming’ step…down?

    Checking on Nain last month, the September mean was 45°F and in September of 1779 it was 45.4°F.

  3. October 22, 2016 11:35 pm

    A good article by Robin. In my previous life I was an engineering hydrologist and I always used the Cummulative Sum (cusum) technique to assess the quality of data before I adopted it for use. Mainly these were for streaflow data which can have sudden shifts in consistency for several reasons including changes in river bed conditions, changes in the discharge rating curves (discharge vs level) or undocumented changes in the station location. Sudden climatic shifts were generally not considered but maybe they should have been.

    Sudden changes in the cusum plots for individual temperature records could be due to changes in the station location, the addition of a nearby building, changes in the instrumentation used or (dare I say) the introduction of jet aircraft to stations located at airports. I would suggest that cusum comparisons be made at a number (say 5 or 6) stations in a region to see if consistent cusum plots are found. If they are consistent, as appears to be the case in SW Greenland, this can be taken of evidence of a regional climatic shift.

    • robinedwards36 permalink
      November 2, 2016 10:15 am

      Hello Brian,
      I have been looking at climate data (ALL Sorts) since about 1992, using cusum methods to try to identify changes and especially sudden changes. They seem to occur very commonly in climate data, and are remarkably consistent over geographically related regions. For example, all eleven Swiss sites that I looked at some years ago displayed remarkably similar cusum patterns, all indicating an abrupt change at 1987, within a month or so of each other. This 1987 change is widespread over Europe and Russia, the onset date drifting a bit later as one goes east. Just recently I have tried looking at some cusums of the residuals from a linear fit to the data, since a “trend” – defined as a linear fit whether or not it is sensible – explains in part (often a small part) of the variance, whether it be meaningful or not! Cusums of the residuals follow the patterns of the un-detrended data, but are even more explicit about some sudden changes. For example the CET monthly data, from which I usually remove the seasonal effect by subtracting the series monthly means from the individual monthly numbers, when addressed via the “residuals route” shows a staggeringly obvious step at Dec 1987. From that date to the present the monthly data have no significant (statistically speaking) change, so the “pause” began in late 1987.
      I could go on for a book’s length about what I’ve found, but won’t!
      Hope to hear from you again. robindottedwardsresiding with talktalkdottnet – if you can decipher this!

Comments are closed.

%d bloggers like this: