The question whether extreme climate events require extreme forcings is assessed for the severe Great Plains drought during May–July 2012. This drought event had a rapid onset, and little indications or early warnings for its sudden emergence existed. The analysis of its origins is based on a dynamical seasonal climate forecast system where states of the ocean, atmosphere, land, sea ice, and atmospheric trace gases were initialized in late April 2012, and an ensemble of forecasts was made. Based on the diagnosis of a spectrum of possible outcomes for precipitation over the Great Plains from this system, it is concluded that the extreme Great Plains drought did not require extreme external forcings and could plausibly have arisen from atmospheric noise alone. Implications for developing early warning system for extreme events in general are also discussed.
 The central premise explored in this study is whether extreme forcing is necessary for extreme climate events, and thus whether an event would have either failed to occur or failed to achieve extreme magnitude if such forcing were lacking. It is sometimes presumed that extreme climate events, such as severe droughts, are the consequence of extreme forcing. Over the tropical latitudes, it is certainly true that extremes in climate variability owe their origin to sea surface temperature (SST) variability associated with El Niño-Southern Oscillation (ENSO) [Hoerling and Kumar, 2002]. The same conclusion, however, cannot be extended to climate variability in extratropical latitudes where contribution of noise associated with random variability in the climate system dominates [Madden, 1976; Kumar and Hoerling, 1995], and the premise that extreme events require extreme external forcings [e.g., see Francis and Varvus, 2012] needs to be assessed with care. The term “forcing” is understood to be a process of subjecting the climate system to a specific and sustained influence. Forcing of drought might involve air-sea interaction processes arising from remote ocean influences (e.g., during extreme phases of ENSO), or land-atmosphere interaction processes arising from antecedent soil moisture influences or land use changes [Cook et al., 2009; Lee et al., 2011]. Other forcings include changes in atmospheric chemical composition resulting from volcanic aerosols, or human-induced greenhouse gas changes. The surmise that extreme events result from extreme forcing, therefore, envisions sustained forcing as being a critical process operating during an event, causing a first-order change in event probability and its magnitude.
 A related supposition is that cause-effect linkages during extremes are sufficiently strong to entail a near-deterministic outcome for a particular climate event, thus implying high predictability in forecasts [Madden, 1976] and high confidence in attribution. These views of forcing and cause-effect relationships are to be distinguished from the interpretation of proximate factors that are also sometimes argued to cause extremes. For instance, it is almost certain that drought results from the absence of precipitation as a key proximate factor, but such knowledge is of no use for predictive understanding (unless reasons for the proximate factors themselves are known and predictable).
 Some clues on how extreme events and extreme forcings may be linked come from the diagnosis of simple dynamical systems. In particular, analysis of atmospheric variability in low-order dynamical models of the atmosphere [e.g., Lorenz, 1963] reveals a wide range of fluctuations. Extreme states occur and can even persist, sometimes transitioning rapidly from one regime to another. The extremes in such models are purely a consequence of nonlinear dynamics occurring in the atmosphere.
 There are several implications that diagnosis of such systems has for addressing our paper's question. First, processes inherent to the atmosphere alone, with no intervention of external forcing, may cause extreme events. Second, the evolution of these atmospheric processes is highly sensitive to initial conditions such that two analyzed states that are only slightly different rapidly diverge. This latter effect is typically viewed as a consequence of the so-called “butterfly effect,” [Hilborn, 2004] and the “mechanism” for extreme events in such nonlinear systems can just be attributable to the (unpredictable) sensitivity to small differences in initial conditions. This paradigm of simple dynamical systems would lead one to propose an alternate hypothesis that extreme climate events do not have to result from extreme forcing, indeed any forcing at all. And, an inability of numerical models to predict an extreme event, even when forcing is present, may reflect a limit to predictability [Madden, 1976] rather than a bias in the model or an error in treatment of the forcing itself. By assessing the influence of possible forcings and of internal variability, we explore two alternate premises for the specific case of the record setting drought that occurred over the central U.S. during summer 2012\.
 Figure 1a shows the May–July 2012 (MJJ 2012) precipitation deficit over the Great Plains (Figure 1a) associated with severe drought during the period. This event had a rapid onset, referred to by some as a “flash drought” and was not anticipated by the official NOAA seasonal drought outlook. The cumulative seasonal rainfall departures for the MJJ 2012 period over the central U.S. was among one of the largest observed within the post-1895 instrumental record. This constituted an abrupt change from conditions during the prior 7 months since the beginning of the hydrologic water year during which time above normal precipitation occurred over most the central Great Plains from October to April 2012 (not shown).
 Several questions can be raised about this extreme event. Was the extreme magnitude of the event related to extreme forcing, for instance the state of global SSTs and/or the cumulative and ongoing effects of global warming? Was the event strongly constrained by such forcings, and thus largely an inevitable consequence of forcings that would have led to high predictability? Or, rather, was the event largely due to random atmospheric variability of the type occurring in the aforementioned low-order dynamical models. The difference in practical consequence between the roles of external forcings versus the role of atmospheric noise is that while the former evolves on a slower time scale and could lead to useful preparedness and hazard mitigation a season in advance, the short lead time associated with the latter cannot provide a comparable useful early warning guidance beyond the short weather predictability time scales.
 In untangling the role of external forcings and atmospheric noise, the current generations of seasonal prediction systems are a powerful tool for analysis. The attributes of the seasonal prediction systems that make them particularly attractive are (a) inclusion of the observed changes in CO2 over time, (b) initialization of observed state of ocean conditions close to the target prediction time, (c) availability of a large set of retrospective forecasts (or hindcasts) to place the forecast anomalies in a historical context, and (d) availability of a large ensemble of forecasts that can be used not only to infer responses to external forcings but also to sample the breadth of possible seasonal mean outcomes that could result as a consequence of atmospheric noise. Using the Climate Forecast System version 2 (CFSv2) [Kumar et al., 2012], the seasonal forecast system operational at the National Centers for Environmental Prediction (NCEP), the role of various factors responsible for MJJ 2012 extreme drought conditions over the Great Plains is analyzed. For this particular example, we demonstrate that the contribution from the atmospheric noise was the foremost cause for rapid onset of drought conditions and that this particular extreme climate event did not arise from exceptional external forcings, and thus, its sudden onset and extreme magnitude could not have been reliably predicted.
 CFSv2 is the operational coupled seasonal prediction system at the NCEP. The real-time seasonal forecasts are complemented by an extensive set of hindcasts over the 1981–2010 period [Kumar et al., 2012]. Hindcasts include four runs for nine target months made every 5 days starting 1 January without considering 29 February in leap years, and in all, there are 73 start dates for every year. All hindcasts are initialized from the observed ocean, atmosphere, and land initial conditions from the Climate Forecast System Reanalysis [Saha et al., 2010]. In its real-time configuration, CFSv2 has four seasonal forecasts initiated each day, and the forecast period covers the next full 9 months. The real-time forecast anomalies are computed relative to the lead-time-dependent climatology obtained from the extensive hindcast data set. Further details about the CFSv2 can be found in Kumar et al.  and Xue et al. .
 The observed precipitation data are from the Climate Prediction Center (CPC) Climate Anomaly Monitoring System-Outgoing longwave radiation Precipitation Index (CAMS-OPI) monthly analysis [Janowiak and Xie, 1999]. The observed SST is from National Climatic Data Center daily high-resolution SST analysis [Reynolds et al., 2007]. The seasonal mean anomalies for both forecast and observation used in this analysis are based on climatology from 1982 to 2010.
 The observed precipitation anomalies for the MJJ 2012 are shown in Figure 1a. By all measures, MJJ 2012 precipitation deficit was unprecedented, and precipitation departures averaged over multistate region (Wyoming, Colorado, Nebraska, Kansas, Missouri, and Iowa) experienced one of the most severe drought conditions in 2012, similar to or exceeding drought conditions in 1988 (M. Hoerling et al., Causes and predictability of the 2012 Great Plains drought, submitted to Bulletin of the American Meteorological Society, 2013). Observed precipitation anomalies over the Great Plains were also accompanied by warmer SSTs in the equatorial tropical Pacific resembling a weak El Niño event (Figure S1a) with the latter providing a potential contributing explanatory factor for the extreme precipitation deficit.
 CFSv2 ensemble mean predictions for MJJ 2012, based on 40 initial conditions between 21 and 30 April (Figure 1b), have a tendency for negative precipitation anomalies over the southern Plains but mostly fail to predict either the location or the intensity of rainfall deficits that occurred over the central Great Plains. The advantage of using ensemble of forecasts from initial conditions at the end of April is that the prediction errors in the SST predictions are smallest for shortest lead time, and as such, the influence of errors in SST forcing is minimized. Indeed, CFSv2 MJJ 2012 SST predictions skillfully captured the observed anomaly pattern and magnitude including warmer SSTs in equatorial Pacific (Figure S1b).
 Several factors could have contributed to the ensemble mean of CFSv2 prediction in Figure 1b: response to SST anomalies; response to change in the CO2; lingering influence of atmospheric initial conditions; and sensitivity to initial soil moisture conditions. It could be argued that the discrepancy between predicted and observed precipitation anomalies may have been due to errors either in the prediction of forcings, or due to errors in responses to the forcings. For example, although the predicted SSTs were close to observed, they were not perfect, and departures from observed SSTs may be responsible for weaker precipitation in the CFSv2. We will revisit the possible role of errors in SST prediction in the context of discussion related to the role of atmospheric noise.
 Discrepancy in the precipitation prediction due to structural uncertainty in the model's precipitation response to the CO2 is unlikely for several reasons. The climatology for all analyses shown herein is the 1982–2010 period, and changes in CO2 amount in 2012 relative to its average over 1982–2010 are not very large. Therefore, even though CFSv2 predictions do include a change in CO2 forcing, its contribution is likely to be small for a bias-corrected prediction system. The argument is further strengthened by previous results that show that a consistent signal of CO2 on precipitation does not emerge until the middle of 21st century [Hu et al., 2012; Deser et al., 2012] or is small in the context of 2012 [Hoerling et al., 2013]. Finally, regarding the influence of the memory of atmospheric and soil moisture initial states on MJJ 2012 ensemble mean prediction, they are more likely to improve prediction skill and not be the cause of discrepancy. Furthermore, the magnitude of these influences on seasonal means is expected to be small because of the rapid divergence of predictions from a narrow cloud of atmospheric initial conditions [Peng et al., 2011; Chen et al., 2010], and due to rapid decay of soil moisture memory [Wang and Kumar, 1998; Guo et al., 2011].
 The arguments given above indicate that it is unlikely that the discrepancy between observations and CFSv2 predictions was due to errors in prediction of the SSTs or to error in responses to either soil moisture or CO2. The conclusions can be strengthened if we can show that the atmospheric noise alone was a major contributor to the MJJ 2012 observed precipitation anomalies, and no external influences were required to generate a precipitation deficit of such exceptional magnitude. This possibility is explored based on the analysis of individual predictions that went into the ensemble mean response.
 To investigate the extent to which the MJJ 2012 U.S. precipitation may have been influenced by the atmospheric noise, we select the four best and the four worst cases out of ensemble of 40 predictions based on anomaly correlation (computed over land point over 20°N–80°N; 175°W–55°W) between precipitation anomaly for observations and individual predictions. The corresponding precipitation anomalies and their difference are shown in Figures 2a to 2c. Important points to note are: there is considerable difference in precipitation anomalies between the four best and worst predictions; the amplitude of precipitation anomaly for the composite of four best cases is at par with the observed precipitation anomaly. Based on additional analysis, we confirmed that the total variability of precipitation predicted by the CFSv2 is not markedly larger than for the observations (not shown), and differences between individual predictions are within the envelope of observed precipitation variability.
 These facts highlight the importance of atmospheric noise in at least contributing to, if not being entirely responsible for, an extreme climate event. In the case of the MJJ 2012 drought, the results indicate that (a) the observed extreme climate anomalies could have resulted from unforced atmospheric variability alone, and thus may have been the outcome of random chance, and (b) that the event may not have been amenable to skillful prediction at longer lead times, and that failures to have anticipated the event by operational drought early warning systems was not due to systemic failure but due to fundamentally low predictability. The importance of atmospheric noise on seasonal mean variability has long been documented based on signal-to-noise analysis and is supportive of this conclusion [Leith, 1973; Madden, 1976; Kumar and Hoerling, 1995; Phelps et al., 2004].
 Revisiting the possible influence of errors in SST prediction, an argument could be made that predicted SSTs among 40 different realizations evolved differently enough to lead to different responses in precipitation. Following this argument, the unique history of observed SSTs and further difference in SSTs between the four best and worst cases could be attributed as the cause for differences in SST with atmospheric noise having smaller contribution. To provide evidence that argues against such an interpretation, we show two additional analyses. Figures 2d to 2f show the composite of predicted SST anomalies for the corresponding four best and four worst cases of precipitation predictions and their differences. The absolute differences in their SST predictions over the equatorial tropical Pacific do not exceed 0.25 K, though larger differences occur in extratropical latitudes over the North Pacific and North Atlantic. These extratropical SST differences are unlikely the cause for the precipitation differences over land among the model composites but are more likely themselves caused by random atmospheric variations. The latter, rather than some strong effect induced by the extratropical SST difference, was also the likely proximate cause for the differences in predicted Great Plains precipitation in Figure 2c. Kumar et al.  discuss this paradigm, and indeed, the differences in atmospheric 200 hPa heights (and inferred surface circulation) (Figure S2) are consistent with the notion that possible cause for SST anomalies was changes in heat flux in the ocean as can be inferred from the low-level circulation associated with the 200 hPa heights in Figure S2. As for the influence of small errors in the prediction of tropical SSTs, previous analysis has shown that differences in equatorial Pacific SST do not have large influence on extratropical response [Kumar and Hoerling, 1997; Barsugli and Sardeshmukh, 2002].
 To further test the possibility that differences in extratropical SSTs may have caused the differences in precipitation, one would need to reduce errors in the SST predictions themselves, and this could be done by reducing the lead time of the prediction. From CFSv2 forecast archives, we construct the seasonal mean of MJJ 2012 by averaging the individual monthly predictions using the shortest lead initializations—prediction of May from the end of April initial conditions, prediction of June from end of May initial conditions, and prediction of July from end of June initial conditions. Although this is not a real seasonal prediction of MJJ 2012 made from initial conditions at the end of April 2012 (as in Figure 1b), the procedure reduces the lead time of “predictions” that went into the construction of seasonal means, thereby reducing errors in predicted SSTs, and allowing a test of the aforementioned hypothesis.
 The four best and worst cases for the constructed seasonal means for MJJ 2012 and corresponding SST predictions are shown in Figure 3. There are still fairly large differences in precipitation anomalies between the composite of four best and worst cases, while the differences in predicted SSTs are now even smaller compared to those in Figure 2. This analysis provides additional evidence to support an interpretation that differences in precipitation in Figure 2 were not due to differences in prediction of SSTs but were due to random atmospheric noise. By extension to the observed extreme drought of 2012, it is then also a plausible conclusion that the extreme event was not the consequence of extreme forcing, but resulted from atmospheric noise.
 From this analysis of a range of possible outcomes for MJJ 2012 precipitation being mainly due to atmospheric noise, it is evident that a much weaker ensemble mean precipitation response (Figure 1b) is to be expected because the noise contribution gets averaged out. A final question remains; however, whether the discrepancy between this ensemble mean precipitation and observations was due to error in the precipitation response to predicted SST, rather than being an indication that it is the random atmospheric noise that drove the observed precipitation and made it differ from the ensemble mean response. To test that conjecture, we analyzed the spatial correlation between ensemble mean and individual realization precipitation predictions over the U.S., and likewise for predicted SSTs between 60°S–60°N. In this so-called “perfect prog” approach, the purpose of the analysis is to determine whether SSTs for individual predictions that are close to the ensemble mean also have their predicted precipitation close to the ensemble mean, or whether large departures in precipitation from the ensemble mean do exist even without differences in SSTs.
 In Figure S3, SST correlations are generally above 0.8 and have little variability from one forecast member to another, while precipitation correlations vary from −0.1 to 0.4. Further, there is no discernible relationship in magnitude of SST and precipitation correlation in so far as higher SST correlations between individual forecasts and the ensemble mean do not correspond to higher precipitation correlation between individual forecasts and the ensemble mean. This indicates that even with SST for individual predictions close of ensemble mean, there could be appreciable range of precipitation outcomes. We conclude therefore that there is a strong possibility that discrepancy between the CFS ensemble mean precipitation response and observations in 2012 was not due to error in precipitation response to SST, but was instead due to random atmospheric noise that mainly drove the observed drought.
 Based on the diagnosis of a spectrum of possible outcomes for precipitation over the Great Plains from a dynamical seasonal forecast system, it is concluded that the extreme Great Plains drought of MJJ 2012 did not require extreme external forcings and could plausibly have arisen from atmospheric noise alone. The broader implications of the analysis were to demonstrate the fundamental role of atmospheric noise in shaping extreme climate events, particularly in extratropical latitudes. This is not to say that the presence of external forcings cannot also influence the magnitude and probabilities of extreme climate events. However, the fundamental dynamics and physical processes of extreme events generally reside in atmospheric processes which themselves may be mostly independent of forcing external to the atmosphere, though in some cases may be modulated by the external forcings such as ENSO or long-term climate change [Hoerling et al., 2013].
 We caution that this is not the case for all the regions over the globe, a point succinctly made by Kumar and Hoerling  in their conceptual depiction of changes in the probability density function of seasonal mean states in tropical and extratropical latitudes [Kumar and Hoerling, 1995, Figure 5]. Over some places in tropical latitudes—the core region associated with ENSO variability, for example—precipitation is mostly controlled by interannual SST variability, and extreme drought and flood events do not occur without appropriate (and extreme) SST forcings [Hoerling and Kumar, 2002]. As a consequence, and given the predictability of SST variations, potential for prediction and an early warning system for climate extremes hold promise. In the extratropical latitudes, however, the same SST forcings change the odds of extreme climate events to a lesser extent. This scenario poses a challenge for developing an appropriate early warning system for extreme climate events, droughts, for example.
 Insightful comments by Scott Weaver, Peitao Peng, Klaus Walter, and an anonymous reviewer are greatly appreciated.
 The Editor thanks Klaus Wolter and an anonymous reviewer for their assistance in evaluating this paper.