Guest Post By Walter Dnes
In continuation of my Temperature Anomaly projections, the following are my May projections, as well as last month’s projections for April, to see how well they fared.
| Data Set | Projected | Actual | Delta |
|---|---|---|---|
| HadCRUT4 2017/04 | +0.606 (incomplete data) | +0.740 | +0.134 |
| HadCRUT4 2017/05 | +0.770 | ||
| GISS 2017/04 | +0.77 | +0.88 | +0.11 |
| GISS 2017/05 | +0.93 | ||
| UAHv6 2017/04 | +0.044 | +0.265 | +0.221 |
| UAHv6 2017/05 | +0.264 | ||
| RSS v3.3 2017/04 | +0.115 | +0.392 | +0.277 |
| RSS v3.3 2017/05 | +0.402 | ||
| RSS v4.0 2017/04 | +0.329 | +0.480 | +0.151 |
| RSS v4.0 2017/05 | +0.497 | ||
| NCEI 2017/04 | +0.7709 | +0.90 | +0.13 |
| NCEI 2017/05 | +0.92 |
The Data Sources
The latest data can be obtained from the following sources
- HadCRUT4 http://www.metoffice.gov.uk/hadobs/hadcrut4/data/current/time_series/HadCRUT.4.5.0.0.monthly_ns_avg.txt
- GISS https://data.giss.nasa.gov/gistemp/tabledata_v3/GLB.Ts+dSST.txt
- UAH http://vortex.nsstc.uah.edu/data/msu/v6.0/tlt/tltglhmam_6.0.txt
- RSS v3.3 http://data.remss.com/msu/monthly_time_series/RSS_Monthly_MSU_AMSU_Channel_TLT_Anomalies_Land_and_Ocean_v03_3.txt
- RSS v4.0 http://data.remss.com/msu/monthly_time_series/RSS_Monthly_MSU_AMSU_Channel_TTT_Anomalies_Land_and_Ocean_v04_0.txt
- NCEI https://www.ncdc.noaa.gov/cag/time-series/global/globe/land_ocean/p12/12/1880-2017.csv Note: has switched to 2 decimal places of precision as of April 2017
The Latest 12 Months
People are already talking about whether or not 2017 will be “the hottest year ever”. The 2016 mean anomaly can be characterized as the “12-month running mean ending in December 2016”. To get an apples-to-apples comparison, May 2016 to April 2017 is used for a 12-month running mean to compare against the year 2016.
The latest 12-month running mean (pseudo-year “9999”, highlighted in blue in the tables below) ranks anywhere from 2nd to 4th, depending on the data set. The May 2017 NCEP/NCAR anomaly is down slightly from May 2016, implying that the 6 May 2017 anomalies will be slightly lower, further cementing the decline of the 12-month running mean. This will make it even harder for 2017 to beat 2016 as the warmest year ever. May marks the 8th consecutive month with NCEP/NCAR global anomaly lower than 12 months ago. However, that could change in June unless the June 2017 value drops below current daily values near the end of May.
The following table ranks the top 10 warmest years for earch surface data set, as well as a pseudo “year 9999” consisting of the latest available 12-month running mean of anomaly data.
| HadCRUT4 | GISS | NCEI | |||
| Year | Anomaly | Year | Anomaly | Year | Anomaly |
| 2016 | +0.775 | 2016 | +0.977 | 2016 | +0.939 |
| 2015 | +0.761 | 9999 | +0.909 | 2015 | +0.903 |
| 9999 | +0.711 | 2015 | +0.858 | 9999 | +0.875 |
| 2014 | +0.576 | 2014 | +0.743 | 2014 | +0.743 |
| 2010 | +0.558 | 2010 | +0.714 | 2010 | +0.703 |
| 2005 | +0.545 | 2005 | +0.692 | 2013 | +0.671 |
| 1998 | +0.537 | 2007 | +0.657 | 2005 | +0.663 |
| 2013 | +0.513 | 2013 | +0.656 | 2009 | +0.641 |
| 2003 | +0.509 | 2009 | +0.643 | 1998 | +0.638 |
| 2009 | +0.506 | 2012 | +0.635 | 2012 | +0.628 |
| 2006 | +0.505 | 1998 | +0.634 | 2006 | +0.618 |
Similarly, for the satellite data sets…
| UAH | RSS v3.3 | RSS v4.0 | |||
| Year | Anomaly | Year | Anomaly | Year | Anomaly |
| 2016 | +0.503 | 2016 | +0.574 | 2016 | +0.781 |
| 1998 | +0.484 | 1998 | +0.550 | 9999 | +0.640 |
| 9999 | +0.360 | 2010 | +0.474 | 1998 | +0.611 |
| 2010 | +0.332 | 9999 | +0.429 | 2010 | +0.558 |
| 2015 | +0.258 | 2015 | +0.383 | 2015 | +0.515 |
| 2002 | +0.217 | 2005 | +0.336 | 2002 | +0.422 |
| 2005 | +0.199 | 2003 | +0.320 | 2014 | +0.414 |
| 2003 | +0.186 | 2002 | +0.316 | 2005 | +0.402 |
| 2014 | +0.176 | 2014 | +0.273 | 2013 | +0.397 |
| 2007 | +0.160 | 2007 | +0.253 | 2003 | +0.386 |
| 2013 | +0.130 | 2001 | +0.247 | 2007 | +0.335 |
The Graphs
The graph immediately below is a plot of recent NCEP/NCAR daily anomalies, versus 1994-2013 base, similar to Nick Stokes’ web page. The second graph is a monthly version, going back to 1997. The trendlines are as follows…
- Black – The longest line with a negative slope in the daily graph goes back to early July, 2015, as noted in the graph legend. On the monthly graph, it’s August 2015. This is near the start of the El Nino, and nothing to write home about. Reaching back to 2005 or earlier would be a good start.
- Green – This is the trendline from a local minimum in the slope around late 2004, early 2005. To even BEGIN to work on a “pause back to 2005”, the anomaly has to drop below the green line.
- Pink – This is the trendline from a local minimum in the slope from mid-2001. Again, the anomaly needs to drop below this line to start working back to a pause to that date.
- Red – The trendline back to a local minimum in the slope from late 1997. Again, the anomaly needs to drop below this line to start working back to a pause to that date.
NCEP/NCAR Daily Anomalies:
https://i1.wp.com/wattsupwiththat.files.wordpress.com/2017/05/daily.png
NCEP/NCAR Monthly Anomalies:
https://i0.wp.com/wattsupwiththat.files.wordpress.com/2017/05/monthly.png
Miscellaneous Notes
At the time of posting, the 6 monthly data sets were available through April 2017. The NCEP/NCAR re-analysis data runs 2 days behind real-time. Therefore, real daily data from April 30th through May 29th is used, and the 30th is assumed to have the same anomaly as the 29th. For RSS and UAH, subsets of global NCEP/NCAR data are used, to match the latitude coverage provided by the satellites.
This month, I’ve switched the land data set projections to use the same algorithm as the satellite data set projection. I.e. the monthly anomaly difference (current month minus previous month) in the NCEP/NCAR subset anomalies is multiplied by the slope() of the data set (versus NCEP/NCAR) for the previous 12 months, and added to the previous month’s anomaly. April actual anomalies for the land sets were more than 0.1 C° above the projections. My previous method was projecting lower May than April values for the land sets, even though NCEP/NCAR anomaly for May is higher than for April. To quote many bad 1950’s B-grade science fiction movies…”That does not compute”.
On the last image it appears that the area below the red trend line is greater than the area above the red trend line, If that is true how could the slope of the line be positive?
It’s not calculated by area. The slope() linear regression uses a mathematical formula that includes the square root of a bunch of squares, etc. If you don’t trust the spreadsheet, but want to “roll your own”, an example FORTRAN implementation is shown at… http://www.pgccphy.net/Linreg/linreg_f90.txt
Adding to WD’s comment, OLS (ordinary least squares) linear trend lines are more influenced by departures from the trend line at the extremities of the data series. This means that the very high value at the left end of the trend is going to have extra “weight” in “pulling” the trend line upward, exaggerating the positive trend. Just eyeballing the data from late 2002 to early 2015 the trend would appear to be essentially flat.
For your April predictions, your deltas are quite large. Have you run any tests to determine if your predictions are any better than a random walk?
I’m not sure what happened. The correlation has fallen off a cliff the last couple of months, even for the surface data sets. I.e. the 12 month correlation versus NCEP/NCAR for the 3 surface data sets
* Mar 2016 to Feb 2017 Hadley 0.8259; GISS 0.9277; NCEI 0.8548
* Apr 2016 to Mar 2017 Hadley 0.5932; GISS 0.8192; NCEI 0.6132
* May 2016 to Apr 2017 Hadley 0.4247; GISS 0.7934; NCEI 0.4252
Always look forward to Walter’s post. On your shorter trends, how are you determining the start dates? I don’t see anything wrong with them, but would like to know how you arrived at them. If discussed in an earlier post, just point to it.
It’s semi-manual, but I make the computer do the hard work first. The spreadsheet calculates the slope() for every day. E.g. when data for May 29 came in, it recalculated slope()
* from Jan 1, 1948 to May 29, 2017
* from Jan 2, 1948 to May 29, 2017
* from Jan 3, 1948 to May 29, 2017
…
* from May 28, 2017 to May 29, 2017
The spreadsheet has a graph of the slope() versus date. I look for noticable local minimums on the graph.
Hovering the mouse pointer over the plot line gives the date+slope() where the pointer is.
Then I go back to the data at the indicated date and check several points before and after that date, until I find the point with the lowest slope() value, in that area.
Walter,
“To get an apples-to-apples comparison”
It’s not clear to me what you numbers for 2015, 2016 etc are here. For app-2-app I would have expected that they would also be running mean to April, but it looks to me more like calendar year figures.
I don’t think it is a particularly good way of estimating the prospects for calendar 2017 vs 2016. The running mean to April 2017 includes the months May-Dec 2016; we know they were relatively cool, and won’t be in he 2017 total. I think a better way is just to look at the YTD average for 2017. We know those months will be in the 2017 average, and in fact YTD 2017 average is about the same as calendar 2016. We don’t know whether the rest of 2017 will be warmer or cooler than YTD, though May was cool in NCAR/NCEP, which dents the prospects for 2017 a little.
Yes, those are annual means for the years (2016, 2015, 2014, etc). We don’t know what the rest of the year will be like. The 12-month comparison gives an idea of what we’re up against. I have 12-month running means plotted at home. HadCRUT and NCEI peaked in August. GISS and RSS3 peaked in September, and UAH peaked in November.
NCEP/NCAR 12-month running mean in September 2016 was the peak and the downward trend started in October, I mentioned in my October 2016 projection https://wattsupwiththat.com/2016/10/31/october-2016-projected-temperature-anomalies-from-ncepncar-data/
In short, every month from August 2015 through September 2016 set a new record for NCEP/NCAR that month. We would need another strong El Nino to beat that. October/November/December weren’t exactly cool either.
When are these statistically indefensible anomalies to the thousandth, and even ten-thousandths, of a degree, going to be replaced by more reasonable numbers?. The surface measurements are taken in tenths of a degree C, and as we all now know from the lengthy discussions we’ve had on the matter, the Law of Large Numbers does not grant greater accuracy to a mean, only greater precision. Plus, large swaths of the Earth’s surface are not represented by actual data at all, but by “interpolations” of existing data that are then used as data, against all scientific and statistical practices.
If those predictions are to have any scientific validity at all, they should be expressed in no more than tenths of a degree, with the error perhaps represented in the hundredths.
When is it going to cool down?
I’m guessing next year.
So both the global NCEP/NCAR anomaly, and the NCEP/NCAR subset for the UAH satellite coverage area go down versus the previous month… and UAH goes up… from +0.265 in April to +0.450 (interim value) in May. I don’t get it.
Walter,
“So both the global NCEP/NCAR anomaly, and the NCEP/NCAR subset for the UAH satellite coverage area go down versus the previous month”
Did it go down? Your graph shows it going up. My calc says it went up 0.06°C. Your surface projections all seem to be going up.
I think it would be a good idea if your posts gave the actual NCEP/NCAR numbers for the last couple of months. It would help explain where the forecasts come from.
But yes, the UAH rise was surprisingly higher.
Sorry, I mis-spoke… what I meant to say was that my projection was for a small drop, but the actual anomaly went up. I looked into it deeper just now. The correlation for UAH versus NCEP/NCAR (UAH subset) is so bad that it’s now slightly NEGATIVE. I.e. as the NCEP/NCAR anomaly goes up, the UAHv6 projection goes down. I may drop projections for the satellite datasets altogether.
Looking at month to month differences, period differences or any chosen period differences seems to me to be inherently very prone to gross but temporary effects. Why not use all available data (assuming of course that the data are viable and believable, which may or may not be valid) and do some statistically unassailable and complete linear analyses, including all appropriate inferential statistics. This is readily doable – I have my own totally reliable software that does all this sort of thing – but is seldom actually done. Not actually a great loss in most instances, because the linear model which is tacitly assumed to be appropriate can (for more than limited time-spans) easily be shown to be completely in appropriate for climate data.
But at least try something useful please.
According to the internet the surface of the earth is 196.9 million square miles. Can we really discuss a temperature anomaly to three decimal places? To my high school educated mind that seems extremely precise (or accurate I’m not sure which).
High precision. Low accuracy.
I would expect successive Summer temps to bounce up and down but mostly down following a strong El Niño. Eventually layered up ocean heat will become depleted until another El Niño hits. System normal.
I see that the actual figures for both UAH (0.45) and RSS (0.482) have been available for a few days.
It’s unusual that there has been no article highlighting these this month……….????