Warm Februaries becoming much more common in the US

The contiguous United States (CONUS) was exceedingly warm east of the Rockies in February 2017, and is notable as being the 27th month in a row with more record highs than lows.

Figure 1 shows monthly mean temperatures up to 6 °C (11°F) above normal. In contrast, the Pacific Northwest was colder than normal. One of the most noteworthy aspects of the persistent February warmth in the U.S. can be seen in the ratio of record high to record low temperatures. According to the National Centers for Environmental Information (NCEI), more than 6,300 record daily highs were tied or broken compared to less than 130 record lows. As a result, February will go down as the 27th month in a row with more record highs than lows. For every low temperature record set there were 49 high temperature records set, the highest such monthly ratio since January 1920.

Temperature anomalies across the US averaged over February 2017
Figure 1. Temperature anomalies averaged over February 2017. ECMWF analysis compared to ERA-interim 1981-2010.

Averaging the monthly mean temperatures over the CONUS, 2017 was found to have the second warmest February on record, behind that of 1954, with a temperature of 4.9 °C (41 °F; Figure 2, based on preliminary data for 2017 from the ECMWF analysis). This was 3.1°C (5.6°F) warmer than the 1981–2010 normal. Averaged over the area east of the Rocky Mountains (approximated by 105° W) it was the warmest February on record. In contrast, in California February temperatures were only just above normal in 2017 after three very warm years. Farther north, in Washington State, February temperatures were 2.5 ºC (4.5 ºF) degrees below normal at −0.8 °C (30.6 ºF).

February temperatures 1895–2016 across the US (NOAA)
Figure 2. February temperatures 1895–2016 (NOAA), the 2017 value estimated from analyses. The green line is a 10-year running mean.

February CONUS temperatures exhibit a warming tendency over the last 120 years (Figure 2). Two questions tied to this apparent warming are: i) is it distinct from changes one would expect from natural weather fluctuations (“detection”)? and ii) if so, can the causes behind the trend be ascertained (“attribution”)? A combined analysis of the historical observations and climate models can help answer these questions.

The analysis of observations provides us with an estimate of how rare the event was, and how its likelihood has changed since the 1900s, a time for which we have good temperature observations but global warming had only just started. The causes of the observed changes can be assessed with climate model experiments with which one can isolate the effects of external factors known to influence the climate: human-caused greenhouse gas and aerosol emissions, and also solar variability and volcanic eruptions.

Trend in the observations

We use the NOAA/NCEI series of CONUS temperatures derived from many thousands of weather stations (Lawrimore et al, 2011). Earlier research has shown that this series is not affected materially by urbanization and siting (Menne et al, 2010, Williams et al 20212); using only rural, well-sited stations actually gives slightly higher trends than the full database. Similar results are found if the analysis is repeated using other surface temperature data (e.g., NASA-GISTEMP, Berkeley Earth and CRU datasets). These temperature anomalies occur mainly in the lowest kilometer of the atmosphere, with satellite observations of the lower troposphere showing anomalies that are about half the magnitude of the near-surface observations. Besides this well-known difference with surface observations, the satellite record is also too short to robustly determine trends in February CONUS temperatures or to assess changes since 1900. We have therefore not included it in this analysis.

The upper tail of February CONUS temperatures is described well by a Normal distribution. We assume that it varies to first order proportional to the global mean temperature, GMST (smoothed), but allow a slower or quicker pace (this method was first used in van Oldenborgh, 2007 and is described fully in van Oldenborgh et al, 2015). The remaining variability is due to the random weather, with virtually no low-frequency variability left over. El Niño – Southern Oscillation (ENSO), the Pacific Decadal Oscillation (PDO) and  Atlantic Decadal Variability (AMV) do not influence the CONUS temperature in February, simplifying the analysis.

This fit, excluding 2017, shows that the CONUS temperature in February has risen a bit faster than the GMST, with a 95% uncertainty range between 0.8 and 2.7 times the global mean temperature. The return period of the 2017 temperature is about 12 years, with an uncertainty range of 5 to 40 years. This means that the chance of getting a temperature this high or higher is around 8% each year, with 95% certainty within the range 2% to 30%. Thus, in the current climate this was not really an exceptional event any more.

Temeprature in February across the contigous US, 1895–2016

February temperatures across the contigous US
Figure 3. Top: February CONUS T2m as a function of the smoothed global mean temperature. The red lines indicate the best fit over the period 1895-2016 (mean and mean plus 1,2 standard deviations), the purple square the value for 2017. Bottom: the same fit as a function of return period (inverse probability) for the climate of 2017 (red) and 1900 (blue). The purple line again indicates the value of 2017. The observations are shown twice, once shifted up with the fitted trend to 2017, once shifted to 1900.

Around 1900 the return period would have been about 160 years, i.e., a probability of 0.6%, with an uncertainty from 0.05% to 1.4% each year. Thus, on the basis of the observed trend we find that the probability has increased by roughly a factor of 13. We have 95% confidence that the increase in likelihood of events like February 2017 is at least a factor of four. However, it is important to note that this increase isn’t all attributable to climate change. For this attribution we rely on climate models.

A similar fit can also give the return period of the 1954 temperature. At that time it was a very unusual situation with a return period of roughly 1 in 200 year (about 0.5% chance). That also means, however, that the odds of finding such an event in 120 years of data are about 50-50, just like throwing a 6 in three throws of dice. The weather maps also confirm that February 1954 had much more unusual weather than February 2017, while the circulation for this year is not as unusual, adding support to the fact that the heating trend has made high temperatures like last month more commonplace than they were in the 1950s.

The same analysis for Washington State, starting in 1900, gives a return period of a February temperature this cold of about once every 30 years, 3% each year. Due to a strong trend, two times the global mean, around 1900 it would have been a very normal year occurring every four years or so.


The next question is what caused the heating trend in the February CONUS observations over the last century. This question cannot be directly assessed using the observational record, with which we can only assess correlation and not causality,  and must be assessed using climate models, in which the various external “forcings” (such as changes in insolation, volcanoes, greenhouse gas concentrations, etc.) can be controlled and their impact understood. We use a collection of coordinated experiments from climate models run at centers across the world under the “5th Coupled Model Intercomparison Project”, or CMIP5, which were assessed in the  IPCC Fifth Assessment Report, the data for which are publicly available. In addition to CMIP5 experiments, two ensembles of atmosphere-only models are used: the UK Met Office HadGEM3-A model at N219 (60km) and the very large ensemble of Weather@Home runs of HadAM3P simulations at N96. Fifteen of the CMIP5 models and the two SST-forced models have been run at least three times with and without human-caused emissions of greenhouse gases and aerosols, allowing us to isolate the effect of the influence of greenhouse gas changes on the likelihood of February warm spells within the climate models.

For the CMIP5 analysis, we first checked which of the 15 models, with the required simulations for the analysis, have a CONUS temperature distribution that is compatible with the observed one, after a constant bias correction, similar to Lewis and Karoly 2013; King et al. 2015. Two models did not pass this test and were removed from the analysis. For the remaining models we counted how many times a temperature anomaly of +3.5 ºC (+6.3ºF) above 1961-1990 was exceeded. This threshold includes the 1954 and 2017 events. We did this three times: in 2007–2027 in the RCP8.5 scenario to estimate the return interval in the climate representative of 2017, in 1901–2005 from the the natural world model experiments (i.e., those without human-caused greenhouse gas and aerosol emissions; NAT), and in a future world around 2050 (2040-2060) assuming the RCP8.5 scenario, which represents continued intensive use of fossil fuels.

Probability Density Functions (PDFs) of CONUS February temperature anomalies in the US
Figure 4. Probability Density Functions (PDFs) of CONUS February temperature anomalies (from a common 1961-1990 baseline) in the natural world ensemble (blue), the current world ensemble (orange) and a future world ensemble (red). The threshold used in the attribution analysis is shown (dashed line).

Historical and projected increases in greenhouse gases lead to increased incidence of warm CONUS Februaries, and decreased incidence of cold CONUS Februaries. The CMIP5 climate model simulations (Figure 4) show an 18-year-return period (95% confidence interval between 12-37 years) for Februaries over the CONUS as warm or warmer than 2017, similar to the observational estimate. Without the historical human-induced increases in greenhouse gases, the “best estimate” of the return period in the models is around 60 years (95% CI between 27-184 years), or a tripling of its probability — the lower bound on the 95% range of the probability change is 1.8x now from past greenhouse gas increases. The model estimate of the greenhouse-driven increase in warm February probability is within the uncertainty range of the observed increase in probability, though  on the lower range of the observational estimates. Around 2050 temperatures like this are projected to be completely normal, occurring approximately every three years on average (95% CI between 2.6-4.5 years).

Of note is that even in the 2050 ensemble, the probability of “cold” winters over the CONUS remains: natural fluctuations in weather and climate are expected to continue generating cool conditions, though at a reduced rate. That is, long-term warming and increased incidence of warm events does not obviate cool events – it just reduces their probability.

The 15 HadGEM3-A historical runs 1960–2015 (Christidis et al, 2013) reproduce the variability and trend of the February CONUS temperatures well, though the model has a systematic 2.3°C cold error in its temperatures, which is adjusted in the analysis,. The simulations with only natural forcings (solar variability and volcanic eruptions) show no trend, showing that the whole trend in the historical simulations from 1960, about 2.1 times the global mean temperature (consistent with the relation estimated in observations), is due to  emissions. The return time of the temperature observed in February 2017 is about 11 years, in agreement with the observed value. In 1900 this would have been around 200 years. The probability of a temperature like the one observed in 2017 or higher has thus increased by a factor of about 20 in this model, with a lower bound of 13 (95% CI, one-sided).

Febrary mean temperatures HadGEM3-A-N216


Trends in February CONUS 2m temperature in HadGEM3-A 1960–2014
Figure 5. Trends in February CONUS 2m temperature in HadGEM3-A 1960–2014. Top: only with natural forcings. Bottom: with all forcings, including greenhouse gases and aerosols.

We analyzed 4898 Weather@Home simulations of February 2017 in current climate conditions forced by seasonal forecast SSTs (Haustein et al, 2016) compared to 3710 possible February temperatures with the anthropogenic signal from atmosphere and SSTs removed. The difference between these two large ensembles reveals a smaller but significant increase in the likelihood of warm Februaries occurring. What is now a 1 in 12 year event would have been a 1 in 22 year event in the world that might have been (risk ratio 1.9 with 95% confidence larger than 1.6) while what would have been a 1 in 160 year event  in the world that might have been would now be a 1 in 62 year event (risk ratio 2.6 with 95% confidence larger than 1.8).

These modeling and observational results are not very surprising given the broad observational, theoretical and modeling background on the connection between increasing greenhouse gases, global and regional temperature changes, and temperature extremes. The character of mechanisms behind the observed rise of the global mean temperature has been studied extensively in the scientific literature, and has been assessed in the Fifth IPCC Assessment Report (AR5, Chapter 10, Bindoff et al, 2013), leading to the conclusion that ‘it is extremely likely that human influence has been the dominant cause of the observed warming since the mid-20th century’. For most of the world, local enhanced likelihood of  warm spells can already be related to this global change, including warming over the contiguous United States in winter.


Figure 6 summarises all the results on the change in probability from the observed trend (blue) and the three model ensembles (red). All results are compatible with a common trend plus natural variability (χ²/dof = 0.9). We therefore computed a weighted average of all results, this is shown in purple: the probability has increased by a factor more than three.

The change in probability of a February CONUS mean temperature as high as the value observed in 2017 or higher
Figure 6. The change in probability of a February CONUS mean temperature as high as the value observed in 2017 or higher. The black lines denote the mean, the bars the 95% confidence intervals. Observations are shown in blue and models in red. The purple bar shows a weighted average of observations and models.


Results from global climate models are inconsistent with the hypothesis that the increase in odds of warm Februaries was caused by natural forcing agents such as solar activity, which has declined since the 1960s, and volcanic eruptions. Meanwhile the model results indicate that past historical increases in greenhouse gases have raised the odds of warm Februaries in the CONUS considerably. The observed trend is compatible with the effects of human-induced emissions of greenhouse gases. Since past and projected future greenhouse gas increases will continue to raise the temperatures, the frequency of winter months like February 2017 should be expected to increase over the coming decades.

Overall, we find that the chances of seeing a February as warm as the one experienced across the Lower 48 has increased more than threefold because of human-caused climate change. The record-warm February of 1954 was, at the time, a very rare event (probability about 0.5% per year) but similar events should now be expected every few years.

Climate change made February US heat 3 x more likely


Bindoff, N.L., P.A. Stott, K.M. AchutaRao, M.R. Allen, N. Gillett, D. Gutzler, K. Hansingo, G. Hegerl, Y. Hu, S. Jain, I.I. Mokhov, J. Overland, J. Perlwitz, R. Sebbari and X. Zhang, 2013. Detection and attribution of climate change: from global to regional supplementary material (pdf, 10.1 MB). In: Climate Change 2013: The Physical Science Basis. Contribution of Working Group I to the Fifth Assessment Report of the Intergovernmental Panel on Climate Change [Stocker, T.F., D. Qin, G.-K. Plattner, M. Tignor, S.K. Allen, J. Boschung, A. Nauels, Y. Xia, V. Bex and P.M. Midgley (eds.)].

Christidis, N., P. A. Stott, A. A. Scaife, A. Arribas, G. S. Jones, D. Copsey, J. R. Knight, and W. J. Tennant. (2013) A new HadGEM3-A-based system for attribution of weather- and climate-related extreme events. J. Climate, 26: 2756–2783 doi: 10.1175/JCLI-D-12-00169.1.

Haustein, K., F. E. L. Otto, P. Uhe, N. Schaller, M. R. Allen, L. Hermanson. N. Christidis, P. McLean and H. Cullen. (2016) Real-time extreme weather event attribution with forecast seasonal SSTs. Environmental Research Letters, 11: 064006. doi: 10.1088/1748-9326/11/6/064006

King, A., G. J. van Oldenborgh, D. J. Karoly, S. C. Lewis and H. Cullen. (2015) Attribution of the record high Central England temperature of 2014 to anthropogenic influences. Environmental Research Letters, 10: 054002. doi: 10.1088/1748-9326/10/5/054002

Lawrimore, J. H., M. J. Menne, B. E. Gleason, C. N. Williams, D. B. Wuertz, R. S. Vose, and J. Rennie. (2011) An overview of the global historical climatology network monthly mean temperature data set, version 3. Journal of Geophysical Research, 116, D19121. doi: 10.1029/2011JD016187.

Lewis, S. C., and D. J. Karoly. (2013) Anthropogenic contributions to Australia’s record summer temperatures of 2013, Geophysical Research Letters, 40: 3705–3709, doi: 10.1002/grl.50673

Menne, M. J., C. N. Williams Jr., and M. A. Palecki. (2010) On the reliability of the U.S. surface temperature record. Journal of Geophysical Research, 115, D11108. doi: 10.1029/2009JD013094

van Oldenborgh, G. J. (2007) How unusual was autumn 2006 in Europe? Climate of the Past, 3: 659-668, doi: 10.5194/cp-3-659-2007

van Oldenborgh, G. J., R. Haarsma, H. de Vries and M. R. Allen. (2015) Cold extremes in North America vs. mild weather in Europe: the winter of 2013–14 in the context of a warming world. Bulletin of the American Meteorological Society, 96: 707–714. doi: 10.1175/BAMS-D-14-00036.1

Williams, C. N., M. J. Menne, and P. W. Thorne. (2012) Benchmarking the performance of pairwise homogenization of surface temperatures in the United States. Journal of Geophysical Research, 117, D05116. doi: 10.1029/2011JD016761