The Marketing of Alt-Data at Temperature.Global

Temperature.Global
(12-month running average, sort of)

A few years ago, I started seeing people post links to a new dataset for global temperatures at a website called temperature.global. There are at least four unique features of this dataset. First, the website claims to update global temperature in near real time. Second, every year in the dataset averages below "normal." Third, there is absolutely no transparency regarding their methodology (those who administer it are anonymous, most of the data is hidden, and much of their methodology is unclear). And fourth, what is disclosed about their methodology reveals that what they are calculating is not the Earth's surface temperature.

Description of Temperature.Global

If you go to the temperature.global website, it becomes almost immediately apparent that this website has little concern for transparency. The website is hosted by Parler (the alternative social network), which if nothing else is pretty odd. There are no names for anyone associated with this published anywhere. We're simply told, "This site was created by professional meteorologists and climatologists with over 25 years experience in surface weather observations." This could be one person with 25 years of experience checking his home weather station or perhaps 25 people with one year experience reading meteorology blogs, or some other combination of people and credentials. I did email them at an email address published on the website. The response I received was signed by "TG" (which is certainly short for temperature.global).

According to the website, they use four data sources: NOAA Global METARs, NOAA One-Minute Observations (OMOs), NBDC Global Buoy Reports, and MADIS Mesonet Data. Once the data is received, it goes through a QC procedure. On the website, they say they use "unadjusted surface temperatures" but don't describe the QC procedure at all. So I sent them an email and asked for their process. TG responded with:

The QA[sic] process looks for several things. First, it looks for extreme outliers (e.g. a temperature of 400 degrees). Next it compares data with its nearest neighbors to look for significant deviations. For example, if Burlington VT is -5F and Middlebury VT is 32F and Barre VT is 33F, then Burlington would get removed. It also compares the current temperature with the previous temperature, and compares the current temperature with the climatological average for that day.

So once the "unadjusted surface temperatures" are received, some adjustments to the data are made, including deleting outliers from the record during the QC process.  The data that passes the QC procedure is then entered into a database, where it undergoes "data functions" to produce a global temperature. In response to one of my questions, TG described these functions:

The data functions are algorithms that create the global mean. It calls the database for the last 12M of data. Some data functions also serve as an API for users to embed the data on their own webpages.

None of the "data functions" involve calculating a global mean surface temperature (GMST), even by the admission of  TG. In order to calculate GMST, you need to use a grid or some other means to handle areas that are sparsely covered with thermometers and those with dense concentrations of thermometers (like cities). You also need to account for station moves and closures and correct for biases introduced by time of observation changes, instrumental changes, and changes in temperature sampling methodology. However, none of this was done. TG says,

The[sic] is no gridding or weighing of data.... There are many organizations that already so[sic] this, like NOAA and NASA. Our project just takes the statistical mean of all available surface data. The intention is to get a different look of[sic] the data without manipulating it at all.

So this is a simple average of thermometer data that passes their QC procedures. It is not a calculation of GMST from instrumental data. The average they produce will be severely biased towards areas with the largest concentrations of thermometers.

What Temperature.global Calls a 12-month Running
Mean Earth's Surface Temperature

The data is presented in several different ways on their site. On the main page, current temperature is displayed as an absolute temperature; at the time I'm writing this, currently: 13.81°C (56.85°F) with a "deviation" of -0.19°C (-0.35°F). The "deviation" appears to be from 14°C. In another portion of the site, there's a list of annual mean temperatures from 2015 which are reported with respect to "normal;" oddly, every year either at normal or below normal. By email I confirmed that what they call "normal" is 14°C. There is also a graph of monthly temperatures on the site, but this is not actually monthly temperatures but a 12-month running average of monthly temperatures which are said to be "compared against the 30 year mean." Given that it's impossible to calculate a 30-year mean from 7 years of data, I asked a question about this as well. Here is TG's response:

The 14C value is the 30 year NOAA average from 1980-2010. NOAA has since changed to the 1990-2020 period but we did not change it to keep it consistent with our previous measurements.

This is a confusing response because NOAA doesn't use a 1980-2010 baseline or a 1990-2020 baseline. NOAA's baseline is the 20th century average. However, UAH once used a 1981-2010 baseline which they did change to a 1991-2020 baseline. So it seems he's confused on this point. But beyond this, NOAA doesn't say any baseline is 14°C. However, NASA does; it estimated that the average temperature of their 1951-1980 baseline was 14°C, and that was accurate to within several tenths of a degree. So I sent a follow up question about this, and I received this response from TG: "Looking into it more, the baseline was from NASA. I couldn't find the exact data source, but I know it came from here: https://data.giss.nasa.gov/gistemp/ and there is an article here: https://www.space.com/17816-earth-temperature.html."

Annual Temperatures are all Below "Normal"

So eventually I was able to clarify that both "normal" and the "30-year mean" are claimed to be 14°C. But neither of these values have anything to do with the temperature.global data. NASA's GISTEMP dataset is an actual GMST dataset, and the 1951-1980 baseline is correctly calculated from their data. It's inappropriate for TG to take a baseline from a GMST dataset and apply it within their own non-GMST temperature record. Their methodology is different (and wrong), their data sources are different, and there's no overlap between NASA's baseline and the temperature.global time series. This leads inevitably to confusion, since they show every year of their annual data to be normal or below normal, and they give the impression that they calculated a 30-year baseline from 7 years of data.

Evaluation of Temperature.Global

We've already seen a lot that's wrong with this time series. It doesn't report what it claims to be - it's not a GMST time series. Since it's a simple average of thermometers, it's biased by areas with the densest concentrations of thermometers. It uses an inappropriate baseline from a legitimate GMST dataset. And there's a nearly complete lack of transparency. But problems continue from here. I decided to check for consistency in the time series. Since the main graph is a 12-month running average, December of every year should be equal to the annual average they report on their website. This happens to be the case every year except for the first two years of the time series:

2015 Avg: -0.54 C     Dec. 2015:  -1.896 C
2016 Avg: -0.27 C     Dec. 2016:  -1.037 C
2017 Avg: -0.26 C     Dec. 2017:  -0.26 C
2018 Avg: -074 C      Dec. 2018:  -0.74 C
2019 Avg: -0.36 C     Dec. 2019:  -0.36 C
2020 Avg: -0.00 C     Dec. 2020: -0.00 C
2021 Avg: -0.11 C     Dec. 2021:  -0.11 C

There really is no excuse for this. If they are presenting an actual 12-month running mean, the value for Dec 2015 and Dec 2016 should equal the reported values for the 2015 and 2015 temperatures, respectively.

Internal Problems with the First Two Years of the Time Series

In one of my emails I asked for the actual monthly data, rather than the 12-month running mean. TG wouldn't give this to me, and he puzzlingly responded as if I asked for access to their database: "We don't give out access to our database, and the data set is so large (terabytes) there's really no way to dump it to a file or anything." But it occurred to me that I should be able to reconstruct the monthly data from the averages provided that the first month of data they have comes from Jan 2015 alone. TG confirmed that this is the case: "We only started the project in 2015, so we do not have any data prior to that. Some people ask why we can't run the algorithms for other years. It's because we are set up to collect data live, so we don't have any code written to go back in to collect and synchronize all prior data."

Since they have no data from 2014, that means that their Jan 2015 value can't be a 12-month running mean. It must be simply their value for the Jan 2015 average. February would then most likely be the average of Jan and Feb; March is the average of Jan, Feb and March, etc. December 2015 would be the first month capable of an actual 12-month average. If that's how TG calculated the running means, then I can reconstruct the original monthly data as follows:

Jan 2015 =  -1.799 C as reported on their website
Feb 2015 =  2*(Feb 2015 running mean) - Jan 2015
Mar 2015 = 3*(Mar 2015 running mean) - Sum (Jan 2015 to Feb 2015)

      ...

Dec 2015 = 12*(Dec 2015 running mean) - Sum (Jan 2015 to Nov 2015)

Every month following Dec 2015 can be reconstructed just like Dec 2015. Assuming I've understood their process correctly, this is what their monthly data would look like.  The blue line is their reported 12-month running mean. The red values are the reconstructed averages, and the thicker faded red line is a calculated 12-month running mean from the reconstructed averages. Basically I'm checking myself to see if I did it right with the faded red line.

Monthly Data (?) for Temperature.Global

The graph shows a definite pattern, especially after the second year. October approaches -30°C below 14°C while March approaches 25°C above 14°C every year. That would mean global temperatures swing by nearly 55°C in just 5 months. I'm calling foul on this. I think whatever is funky about the first two years of data that makes the Dec values for 2015 and 2016 not equal to the annual average is messing up my reconstruction. I also think this time series is completely worthless no matter how you look at it. It's not even consistent with itself. I have sent the above graph to TG for them to confirm or correct my reconstruction, and I asked more questions about the discrepancies in the first two years. TG's initial response is less than satisfying. "2015 was a little different because we didn't have a year's worth of data at that point. I'll have to investigate why Dec 2016 isn't the posted average for 2016 (it should be, as you noted). The monthly data point should be available on the graph, but if not, I can send you a table." I responded asking for the table and clarifying that Dec 2015 should also be equal to the posted average for 2015. As of the point of writing I haven't heard back from them, but I'll amend this post with further details should I receive any response from TG.

Update: 2/25/2022

It's now been a week since TG has responded to any of my emails. He was initially very responsive, sometimes responding in a matter of hours. But since I asked for the table of monthly data, he has not responded.