The frequency of cold months in the 21st century is studied using the CMIP3 ensemble of climate model simulations, using month-, location- and model-specific threshold temperatures derived from the simulated 20th century climate. Unsurprisingly, cold months are projected to become less common, but not non-existent, under continued global warming. As a multi-model mean over the global land area excluding Antarctica and under the SRES A1B scenario, 14% of the months during the years 2011–2050 are simulated to be colder than the 20th century median for the same month, 1.3% colder than the 10th percentile, and 0.1% record cold. The geographic and seasonal variations in the frequency of cold months are strongly modulated by variations in the magnitude of interannual variability. Thus, for example, cold months are most infrequently simulated over the tropical oceans where the variability is smallest, not over the Arctic where the warming is largest.
 Despite the observed warming of the global climate, periods of cold weather still occur. As a recent example, December 2010 was the coldest December during the past century in parts of northwestern Europe, including Great Britain [Met Office, 2011]. Although such cold periods sometimes seem incompatible with the concept of global warming to the general public, they are not. Due to the variability of the atmospheric and oceanic circulation, occasional below-average or, in rare cases, record cold temperatures, are expected to occur even when the global mean temperature is increasing. Yet, an important question is, how often? Is the recently observed occurrence rate of cold weather consistent with model simulations of ongoing climate change? How fast (if at all) is the frequency of cold periods likely to decrease in the future?
 Model simulations support a continued trend toward generally milder and less frequent cold extremes in the future. Kharin et al.  found a globally averaged 2.3°C increase in the 20-year return level of the lowest yearly minimum temperatures from 1981–2000 to 2046–2065, as an ensemble mean over 12 CMIP3 (Coupled Model Intercomparison Project 3 [Meehl et al., 2007a]) models run under the Special Report on Emissions Scenarios (SRES) A1B scenario [Nakićenović et al., 2000]. In large parts of the world, the statistical waiting time for temperatures corresponding to the 20-year return level in 1981–2000 was found to increase to several centuries by the mid-21st century. Other studies based on the CMIP3 ensemble qualitatively support these findings, but have highlighted substantial variations between different models and areas [Vavrus et al., 2006; Kodra et al., 2011]. For example, Vavrus et al.  found cold air outbreaks (defined using temporally fixed temperature thresholds based on the late 20th century climatology) to decrease by 50–100% in frequency in most of the Northern Hemisphere during the 21st century. Still, due to changes in atmospheric and oceanic circulation, some limited areas had more cold air outbreaks in the future in some of the models.
 In apparent conflict with other model studies, Petoukhov and Semenov  found that a decrease in sea ice in the Barents and Kara seas might induce circulation changes that would make cold winter extremes more common in parts of Eurasia and northern North America. However, their study only addressed the climatic response to changing sea ice conditions, neglecting changes in the atmospheric composition and sea surface temperatures. Furthermore, the apparent response in atmospheric circulation and cold extremes was non-linear, depending strongly on month and the magnitude of the sea ice decrease.
 Most of the studies discussed above have focussed on conditions in middle or late 21st century. From the practical point of view, the occurrence of cold weather in the next few decades is at least of equal interest. Furthermore, while cold extremes in winter play a special role, anomalously cool weather in other seasons also deserves more attention than it has received this far.
 Here, we complement earlier studies by systematically analysing the frequency of cold months in the CMIP3 models. We consider three month-, location- and model-specific thresholds of coldness, all defined relative to the simulated 20th century climate. Months with mean temperature below the median of the same month in 1901–2000 are termed cool, those below the 10th percentile as very cold and those colder than any month in 1901–2000 as record cold. The term cold months refers to these categories collectively.
 Specifically, we are motivated by the following questions:
 1. How is the frequency of cold months likely to change with time? How does this vary with the time of the year? When do differences between emissions scenarios become important?
 2. How large is the projection uncertainty in the occurrence rate of cold months, as indicated by the variation among the CMIP3 simulations?
 3. Is the occurrence rate of cold months predicted well by changes in the long-term mean temperature, or are changes in interannual variability also important?
 4. Has the frequency of cold months during the early 21st century been consistent with model projections?
2. Data Sets
 Data from 24 coupled atmosphere-ocean models in the CMIP3 intercomparison are used (Table S1 in the auxiliary material). For each model, one or more continuous time series covering the years 1901–2098 were produced by concatenating the 20th century simulations (20C3M) with 21st century simulations based on SRES emissions scenarios. Most of our analysis focuses on the SRES A1B scenario, using data for all 24 models. However, we also study the scenario sensitivity of our findings, using the subset of 17 models for which all of the SRES A1B, B1 and A2 scenarios are available. Except when studying the role of internal variability (Section 2 of auxiliary material), only one realization of each scenario is used for each model. Prior to the analysis, all model data were regridded to a regular 2.5° × 2.5° latitude-longitude grid using the nearest-neighbor method.
 For evaluating the simulated frequency of cold months in the decade 2001–2010 against observations, the University of East Anglia Climate Research Unit temperature analysis version CRU TS3.0 (http://badc.nerc.ac.uk) [see also Mitchell and Jones, 2005] was merged with the ERA-Interim reanalysis [Dee et al., 2011]. The former covers the years 1901–2005 (over land north of 60°S), the latter 1979–2010. Quasi-homogeneous temperature time series for the full period 1901–2010 were created by extending the CRU analysis by five years with the ERA-Interim data, after first adjusting the latter for inter-dataset differences in mean value and interannual standard deviation during the common period 1979–2005.
3.1. Cold Months During the Years 2011–2050, A1B Scenario
 Statistics of the simulated number of cool, very cold and record cold months during the period 2011–2050 under the A1B scenario are shown in Figure 1. Under 20th century climatic conditions, about 240, 48 and 5 cases per 480 months in these three categories would be expected. The actual multi-model mean numbers in 2011–2050 are much lower. Area means over land north of 60°S (representing broadly the inhabited part of the world, and the area covered by the CRU data) are 68 cool, 6 very cold and 0.4 record-cold months during the 40-year period.
 The number of cold months varies geographically. Some aspects of this variation parallel the distribution of the multi-model annual mean temperature change [Meehl et al., 2007b, Figure 10.8]. For example, the area of minimum warming over the northern North Atlantic stands out with the largest number of months in all three categories.
 However, in apparent conflict with the time mean warming, the number of cold months generally increases from low to high latitudes. Similarly, despite smaller warming over sea than land areas, there are at most latitudes fewer cold months over the oceans than over the continents. This is because the impact of the time mean temperature change is strongly modulated by geographic differences in interannual variability [Ruokolainen and Räisänen, 2009; Mahlstein et al., 2011]. Over the extratropical continents, where the variability is strong, a substantial number of cold months will occur even after a moderately large warming. Over low-latitude oceans, where the variability is much smaller, even a modest warming will make cold months much less frequent. A case in the point is the eastern tropical Pacific with its El Niño – Southern Oscillation variability, where a larger number of cold months occur in 2011–2050 than elsewhere over the tropical oceans.
 Differences in interannual variability also manifest themselves in Figure 2. Over much of the Northern Hemisphere extratropical continents, cool months are more common in winter and spring than in summer. In high latitudes, in particular, the time mean warming in winter exceeds that in summer [Meehl et al., 2007b, Figure 10.9], but this is overcompensated by larger interannual variability in winter than summer temperatures. However, there are also regions where seasonal differences in time mean warming dominate over differences in interannual variability. For example, fewer cold months are simulated over the Arctic Ocean in winter (when the warming is largest) than in summer (when the variability is smallest). Similar conclusions hold for very cold and record cold months, although their absolute numbers are much lower.
 The number of cold months varies widely between the 24 models (Figures 1, left and 1, right). The average grid-box-scale range for cool months exceeds a factor of six (as averaged over land north of 60°S, from 20 to 130), with even larger relative variation for very cold months. In wide areas mainly over low-latitude oceans, all months in 2011–2050 are warmer than their 20th century median in one or more models. At the other extreme, in some high-latitude sea areas a majority of the months in 2010–2050 are cool or (in the northern North Atlantic) even record cold in at least one model. As a rule, the annual mean temperature decreases in the same areas and models, due to either forced or internally generated changes in ocean circulation [Intergovernmental Panel on Climate Change, 2007].
 For an ensemble of model simulations run under the same emissions scenario, differences in simulated climate change result from two factors: directly from differences in models and in the implementation of the forcing [Meehl et al., 2007b, Table 10.1], and from different realizations of simulated internal variability. However, when cold month frequency in the full period 2011–2050 is considered, internal variability appears to be a secondary source of variance between the model simulations, even though its importance increases towards the most extreme events (Section 2 of auxiliary material).
3.2. Time and Emissions Scenario Dependence of Cold Month Frequency
 The simulated frequency of cold months decreases gradually with time (Figure 3). For the A1B scenario, the multi-model mean frequency of cool months over land north of 60°S is 22% in 2011–2020, 8% in 2041–2050 and only 3% in 2091–2098. The reduction in very cold and record-cold months is even steeper, exceeding an order of magnitude from the beginning to the end of the century.
 The results for the B1 and A2 scenarios parallel those for A1B in the early 21st century. In the end of the century, cold months in all three categories are least frequent for the A2 and most frequent for the B1 scenario, B1 deviating more from A1B than A2 does. The multi-model mean frequency of cool months under the B1 scenario is still almost 7% in 2091–2098. This exceeds the 24-model maximum for the A1B scenario, and the same holds for very cold but not record cold months. B1 already has the largest frequency of cold months in the mid-21st century when, however, A2 also features slightly more of them than A1B. These differences are consistent with the corresponding inter-scenario differences in the simulated global mean warming [Meehl et al., 2007b, Figure 10.4].
 The observed (CRU + ERA-Interim) frequencies of cool, very cold and record-cold months in the decade 2001–2010 were 22%, 3.0% and 0.24%, respectively (similar numbers are found for 2001–2005 and 2006–2010 separately, despite the switch between the two data sets). These values agree closely with the corresponding multi-model mean results (purple triangles in Figure 3). This supports the use of the multi-model mean projection as a best estimate for the future, particularly as intermodel variations in the area-averaged cold month frequencies turn out to be strongly correlated between the years 2001–2010 and later periods (Section 3 of auxiliary material).
3.3. Are Changes in Interannual Variability Important?
 Although both geographical and seasonal variations in interannual variability are important for the simulated frequency of cold months, changes in interannual variability turn out to be relatively unimportant. Figure 4a shows a reconstruction for the multi-model mean number of very cold months during the years 2011–2050, derived using a method that takes into account the long-term mean temperature change but assumes unchanged interannual variability (Section 4 of auxiliary material). The spatial correlation between the reconstruction and the actual multi-model mean shown in Figure 1 is 0.98. The corresponding correlations for cool and record cold months are 0.99 and 0.96, respectively (maps not shown).
 The actually simulated frequency of very cold months slightly exceeds the reconstruction over most low-to-mid-latitude areas but is generally lower over the high-latitude oceans (Figure 4b). These features qualitatively mimic simulated changes in interannual temperature variability, as shown for an earlier model generation by Räisänen : where the mean temperature increases, increases (decreases) in variability always tend to make cold months more (less) frequent. Yet, the systematic difference between the actual simulations and the reconstruction is modest, at least as averaged over the 12 months and the 24 models. In individual model simulations, changes in variability may play a larger role in some regions and seasons [e.g., Schär et al., 2004].
 In contrast with our results, Ballester et al.  found changes in both the standard deviation and the skewness of the distribution to be important for changes in extreme daily temperatures in an ensemble of regional climate change simulations for Europe. Two factors probably contribute to this difference. First, extreme daily temperatures are more prone to be affected by changes in variability than extreme monthly temperatures are, because variability on shorter time scales is larger and has therefore more room to change with changing climate. Second, Ballester et al.  defined their mean change as the change in the annual mean temperature, so including changes in the seasonal cycle within the change in variability.
 Our main conclusion is unsurprising: individual cold months are to be expected even in future decades, but their frequency is projected to decrease as the greenhouse gas induced warming of the global climate continues. As a multi-model mean over land excluding Antarctica, 68 (14% of 480) months colder than the median for 1901–2000 are simulated during the period 2011–2050 under the SRES A1B scenario. The corresponding numbers for months below the 10th percentile and record cold months are 6 (1.3%) and 0.4 (0.1%), respectively. These averages hide large differences between the 24 models and between different parts of the world. The frequency of cold months in the early 21st century is nearly the same for the B1, A1B and A2 scenarios, but the differences between the scenarios grow much larger towards the end of the century.
 The simulated changes in cold month frequency are to a good approximation explained by the long-term mean warming, with changes in interannual variability playing only a secondary role. Nevertheless, differences in interannual variability are important for the geographical and seasonal distribution of anomalously cold months. Geographically, such months are most infrequently simulated over the tropical oceans where the variability is smallest, not in high northern latitudes where the warming is largest. Seasonally, Northern Hemisphere high-latitude continents are projected to experience the largest number of cold months in winter and spring, despite a maximum of warming in winter. Although perhaps not intuitive, these findings are consistent with earlier research on the signal-to-noise properties of greenhouse gas induced warming [Räisänen and Ruokolainen, 2008; Ruokolainen and Räisänen, 2009; Mahlstein et al., 2011].
 In short, cold months as defined against 20th century climate are still expected to occur in the future, although gradually more seldom. In rare cases, even individual record cold months are likely to occur, this being not inconsistent with projections of continued global warming.
 We acknowledge the modeling groups, the Program for Climate Model Diagnosis and Intercomparison and the WCRP's Working Group on Coupled Modelling for making available the WCRP CMIP3 multi-model dataset. Support of this dataset is provided by the Office of Science, U.S. Department of Energy. This research is part the Academy of Finland project 127239 and of the SETUKLIM project.
 The Editor thanks the two anonymous reviewers for their assistance in evaluating this paper.