1. Introduction
Human forecasts of tropical cyclones (TCs) rely greatly on numerical models. Each model used by forecasters at the National Hurricane Center (NHC) has unique strengths and weaknesses. Research has investigated the skill of these models with respect to track and intensity, with the assumption that a TC already exists (e.g., Goerss 2000; Goerss et al. 2004; Sampson et al. 2008). Several studies have investigated global models' ability to predict TC genesis at various lead times in the western North Pacific (WNP) basin using various models and periods of study (e.g., Briegel and Frank 1997; Chan and Kwok 1999; Cheung and Elsberry 2002; Elsberry et al. 2009, 2010, 2011; Tsai et al. 2011). However, insufficient research has focused on the skill of model forecasts of TC genesis in the North Atlantic (NATL) basin.
The goal of this study is to quantify the accuracy of model-indicated TC genesis through 96 h in the NATL basin. Specifically, we are interested in analyzing the relative performance of multiple global models. Thus, we compare results from Environment Canada's Global Environmental Multiscale Model (CMC; Côté et al. 1998a,b), the European Centre for Medium-Range Weather Forecasts global model (ECMWF 2012), the National Centers for Environmental Prediction (NCEP) Global Forecast System (GFS; Kanamitsu 1989), the Navy Operational Global Atmospheric Prediction System (NOGAPS; Rosmond 1992), and the Met Office global model (UKMET; Cullen 1993). These forecasts are verified against the genesis of best-track (BT; Jarvinen et al. 1984; McAdie et al. 2009) TCs.
Each year approximately 80–90 TCs develop globally (Frank and Young 2007). This paper focuses on the NATL given its relevance to the NHC, but the algorithm employed can be applied globally in the future to assist other operational centers. Tory and Frank (2010) list five conditions that are necessary for TC genesis based on a revision of Gray's (1968) first global TC genesis climatology:
“1) Sea surface temperatures (SST) above 26.5°–27°C coupled with a relatively deep oceanic mixed layer (~50 m) [it should be noted that recent genesis potential indices (e.g., Emanuel and Nolan 2004; Camargo et al. 2007) have abandoned SST in favor of the more theoretically justified potential intensity (Emanuel 1988)];
2) a deep surface-based layer of conditional instability;
3) enhanced values of cyclonic low-level absolute vorticity;
4) organized deep convection in an area with large-scale mean ascent and high midlevel humidity; and
5) weak to moderate (preferably easterly) vertical wind shear.”
The entirety of the processes that cause TC genesis still is largely unknown. This is evident by the number of TC genesis theories that have been proposed, and the disagreement that exists among them. One of the oldest theories is Charney and Eliassen's (1964) conditional instability of the second kind (CISK), which since has been largely disproved as a mechanism for TC development (Craig and Gray 1996). Emanuel (1986) proposed wind-induced surface heat exchange (WISHE), which, while a more accepted theory for steady-state TCs, has deficiencies for explaining genesis (Craig and Gray 1996; Montgomery et al. 2009). Ritchie and Holland (1997) and Simpson et al. (1997) were the initial proponents of the “top down” TC genesis theory. There also is a “bottom up” theory proposed by Montgomery and Enagonio (1998) and Enagonio and Montgomery (2001). A more recent TC genesis theory is proposed by Dunkerton et al. (2008). Several field experiments [e.g., Pre-Depression Investigation of Cloud-Systems in the Tropics (PREDICT; Montgomery et al. 2012), Genesis and Rapid Intensification Processes (GRIP; Braun et al. 2013), Intensity Forecasting Experiment (IFEX; Rogers et al. 2006)] have been conducted to evaluate these theories, but results have not yet yielded an entire understanding of the processes that cause TC genesis.
The operational models are not bound by the above-mentioned theories or the conditions stated to be necessary for TC development; thus, they may not predict development based on physically accepted reasons. Also, due to grid spacing and computational limitations, the models are not able to resolve all of the atmospheric processes governed by the full Navier–Stokes equations. For example, convection, planetary boundary layer processes, and microphysical processes all are parameterized. Therefore, one cannot expect the models to fully resolve all of the processes that are necessary for TC genesis. Despite the above caveats, the models commonly indicate TC-like development in the forecast fields (e.g., Pasch et al. 2006, 2008).
While much time and resources have been devoted to improving model-derived TC track and intensity forecasts, the literature contains relatively few assessments of the models' ability to predict TC genesis in the NATL basin or determine what model enhancements should be made to improve that prediction. Several modes of development exist in the NATL basin (McTaggart-Cowan et al. 2008): nonbaroclinic (40%), low-level baroclinic (13%), transient-trough interaction (16%), trough induced (3%), weak tropical transition (13%), and strong tropical transition (15%). In addition, the significance of factors driving development varies from month to month. For example, TCs forming from African easterly waves (AEWs) are most prevalent from August to mid-September (Gray 1968). Because of this variability, “the North Atlantic is the most complicated development region” (Gray 1968).
Early investigations of model TC genesis forecasts in the NATL indicated that while the models sometimes accurately predict TC genesis, they frequently generate TCs that do not develop (Beven 1999). Conversely, a TC sometimes develops that has not been forecast in the models. The relative performance of three global models (GFS, NOGAPS, UKMET) in predicting TC genesis was analyzed during the 2005 NATL TC season (Pasch et al. 2006). Results showed that the GFS had the greatest probability of detection (POD), but also the greatest number of false alarms (FAs). Conversely, NOGAPS exhibited both the smallest POD and the smallest number of FAs. Pasch et al. (2008) indicated that the accuracy of TC genesis forecasts varies greatly from one TC to another. During 2007, the genesis of Dean (an eventual category 5 hurricane) was well predicted by the GFS several days in advance. However, several weeks later, the GFS failed to predict Felix (another eventual category 5 hurricane).
The current study expands upon these previous works by analyzing additional models and including additional TC seasons in the dataset to provide a more robust climatology of model-indicated TC genesis forecasts. It is important to define metrics for the forecast cases' varying degrees of success. In general the model-indicated forecasts can be verified by whether or not TC genesis actually occurred. But, there are cases where some, but not all, aspects of the forecast are correct. So, while it is important to analyze successful forecasts, we must also examine partially successful forecasts, as well as FA cases and cases when a TC formed in reality but was missed by the global models. Hypotheses regarding the reasons for the analyzed relative performance of the models are beyond the scope of this paper.
Section 2 describes the datasets and methods that were used to define, identify, track, and classify each model-indicated TC as a hit, an FA, or an early genesis (EG) or a late genesis (LG) event. The results (section 3) include metrics for evaluating the models' overall performance as well as geographic and temporal breakdowns of that performance. The summary and conclusions are presented in section 4.
2. Methodology
a. Data
A robust analysis of TC genesis forecasts required a sufficiently large dataset. A local archive of global operational model data included output from the CMC, GFS, NOGAPS, and UKMET global models from 2004 to 2011. ECMWF model output was provided by The Observing System Research and Predictability Experiment (THORPEX) Interactive Grand Global Ensemble (TIGGE) data archive (ECMWF portal) and was available for the 2007–11 seasons. Table 1 summarizes features of each model (Bélair et al. 2009; NHC 2011; ECMWF 2012; Environmental Modeling Center 2012; NRL 2012; COMET 2012; J. Heming 2012, personal communication). Appendix A describes selected upgrades to each model during the period of study. Since data were unavailable for some periods, Table 2 shows start and end dates of the dataset by season and which TCs and models are not included during the available time period. The period of study includes some of the most active NATL hurricane seasons on record. Our dataset contains 135 BT TC genesis events, with Fig. 1 showing their geographic distribution. TC genesis locations are defined as the first time that NHC designated the cyclone as a tropical depression (TD) or tropical storm (TS). For example, the location of tropical cyclogenesis for Vince (2005) is different than the location when advisories were initiated because Vince initially was a subtropical cyclone. Storms that were subtropical throughout their entire life cycle were not considered.
Summary of model characteristics (Bélair et al. 2009; NHC 2011; ECMWF 2012; Environmental Modeling Center 2012; NRL 2012; COMET 2012; J. Heming 2012, personal communication).
Date ranges of model output in the dataset. Storms and models not included in the dataset are listed. The dataset contains 135 storms.
NHC BT genesis positions for the 2004–11 seasons, where genesis is defined as the first entry of a tropical depression or tropical storm.
Citation: Weather and Forecasting 28, 6; 10.1175/WAF-D-13-00008.1
b. Definition of model TC
An important component of this research was to define a model-simulated TC. Given the number of processes that are parameterized in the models due to resolution and computational limitations and the limited number of in situ observations that are assimilated into the models, it is unreasonable to expect that a model-indicated TD will exhibit the same characteristics as a TD in reality. All previous studies on this topic have defined model-indicated TCs differently (e.g., Briegel and Frank 1997; Chan and Kwok 1999; Cheung and Elsberry 2002; Pasch et al. 2006, 2008). Additionally, the algorithm for TC detection and tracking that is most used operationally is that of Marchok (2002). Walsh et al. (2007) showed that differences in model resolution (which exist in this study) lead to different thresholds for parameters that define model-indicated TCs. Although we used the previous definitions as a starting point, we ultimately defined our own criteria (described below) to address as many findings from the previous literature as possible. Thus, all of the experiments initially were required to meet the following criteria:
a relative minimum in mean sea level pressure (MSLP),
a relative maximum in 850-hPa relative vorticity (RV) within 2° latitude and longitude of the MSLP minimum (hereafter, “latitude and longitude” will be omitted when referring to distance in degrees),
a relative maximum in 250–850-hPa thickness within 2° of the MSLP minimum,
the 925-hPa wind speed must exceed the model-specific wind threshold (explained below) at any point within 5° of the MSLP minimum, and
criteria 1–4 must be met for at least 24 consecutive forecast hours.
Criteria 1–3 are similar to the TC definition of Cheung and Elsberry (2002), where criterion 3 is the proxy for a warm core. Criterion 4 addresses the objective wind threshold, and criterion 5 addresses the temporal threshold suggested by Walsh et al. (2007). With the above as a baseline, additional requirements were investigated. For example, we experimented with requiring not only a relative minimum in MSLP, but also at least one closed isobar at various intervals in an attempt to filter out weak, broad areas of low pressure that commonly exist in the tropics, but are not TCs. Also, instead of requiring only relative maxima in criteria 2 and 3, we experimented with different threshold percentiles [e.g., lowest tercile, as in Pasch et al. (2006)].
To investigate these thresholds, we considered how observed TDs are represented in the model environment. We compiled all BT TD entries during the 2008–10 TC seasons, which were the most recent complete seasons at the time of this study (only at the beginning of the storm's life cycle). Then, the corresponding model analysis field at the time and location of each BT TD entry was selected. If criteria 1–3 were met, the values of 850-hPa RV and 250–850-hPa thickness maxima were recorded. Similarly, if criteria 1–3 were met, the maximum 925-hPa wind speed within 5° of the MSLP minimum was recorded. These values allowed us to pick thresholds based on a given percentile (e.g., the lowest tercile value of recorded 850-hPa RV maxima was used as the 850-hPa RV threshold value that must be exceeded to be considered a TC in the model). We recorded these values separately for each model. Although we use the same percentile for every model, the actual threshold values vary from one model to another (Table 3).
Lowest tercile threshold values for each model.
The TCs then were tracked throughout the progression of each model cycle. If the aforementioned criteria 1–4 were met in consecutive forecast hours (e.g., forecast hours 12 and 18) of the same model cycle and the MSLP centers were within 3° of each other, the features were considered to be the same TC (see Fig. 2). If a TC was tracked for at least 24 h (e.g., existed at forecast hours 12, 18, 24, 30, and 36), it satisfied criterion 5.
The blue circles indicate instances where TC criteria 1–4 were met. The blue box is a 3° buffer around the position at forecast hour 12. Note that the position at forecast hour 18 is on or within the 3° buffer. Thus, these two points are considered to be the same TC. A new buffer is then drawn around the location at forecast hour 18 (green box), and a search is done for any positions on or within the buffer at forecast hour 24. This process is repeated until no points are found on or within the buffer.
Citation: Weather and Forecasting 28, 6; 10.1175/WAF-D-13-00008.1
Since one of our primary goals was to provide operational forecasters with information about predicting TC genesis, we conducted a series of experiments with different sets of criteria and threshold values and determined their hit, miss, and FA rates during selected TC seasons. Table 4 provides a summary of the experiments conducted. We tested various thresholds (e.g., 10th percentile, lowest tercile, etc.) and various closed isobar requirements (e.g., 2- versus 4-hPa closed isobar). Also, we tested the criteria at different levels (e.g., 850- versus 700-hPa RV maximum). The criteria and threshold values for each model that optimized the success rate and number of storms detected were used to define a TC (Fig. 3):
a relative minimum in MSLP with at least one closed isobar at a 2-hPa interval;
a relative maximum in 850-hPa RV within 2° of the MSLP minimum; this maximum must exceed the lowest tercile value of all model-indicated TD analyses that correspond to BT TDs;
a relative maximum in 250–850-hPa thickness within 2° of the MSLP minimum; this maximum must exceed the lowest tercile value of all model-indicated TD analyses that correspond to BT TDs;
the wind speed at 925 hPa must exceed the lowest tercile value of all model-indicated TD analyses that correspond to BT TDs at any point within 5° of the MSLP minimum; and
criteria 1–4 must be met for at least 24 consecutive forecast hours.
List of experiments conducted to determine the model-indicated TC criteria. Min/max refers to relative minimum/maximum. Percentiles indicate a threshold value that must be exceeded in the model. The setup chosen for this study is set in boldface.
Schematic of the set of criteria used to define a TC in the model. All four criteria must be met for at least 24 consecutive forecast hours. The dashed box represents a 2° buffer around the MSLP minimum and the solid box represents a 5° buffer around the MSLP minimum.
Citation: Weather and Forecasting 28, 6; 10.1175/WAF-D-13-00008.1
The requirement of at least one closed isobar at a 2-hPa interval removed many weak relative MSLP minima that were equatorward of 10°N, in or near the intertropical convergence zone (ITCZ). It also removed relative MSLP minima that were not cyclones, but merely broad areas of relatively low pressure between two anticyclones. The requirement that the 250–850-hPa thickness maximum exceed the lowest tercile threshold removed many cyclones in the midlatitudes that had smaller thicknesses and likely were not purely tropical.
c. Classification of cases
Once all model-indicated TCs were identified, the cases were filtered and classified. All cases with a model-indicated TC at the analysis time (forecast hour 0) were removed because we are interested only in forecasts of model genesis (i.e., genesis occurring at or after forecast hour 6). The TCs then were classified as a hit, FA, EG, or LG event based on the following definitions:
Hit—a model TC exhibits genesis within 24 h of the BT genesis time and is located within 5° of the BT genesis location;
Early genesis (EG)—TC genesis is forecast in the model; it occurs at the same time as an existing combined automated response to query (CARQ) entry in the Automated Tropical Cyclone Forecasting (ATCF) system A-deck dataset1 (Sampson and Schrader 2000) and is within 5° of the CARQ entry; however, it is not the genesis position of a TC; these are cases when the model predicts genesis for a disturbance that ultimately becomes a BT TC, but the model predicts genesis too early;
Late genesis (LG)—TC genesis is forecast in the model; it occurs at the same time as an existing BT position and is within 5° of the BT location; however, it is not the genesis position of a TC; instead, it is a BT position of a TC that already has formed; these cases are considered forecasts when the model predicts the genesis of an actual TC too late (see Fig. 4); and
Fig. 4. An example of an LG event for TC Colin in 2010. The model-indicated genesis time matches a BT entry time, and the location of model-indicated genesis is within the 5° buffer of the BT entry location. But, the verifying BT position is a postgenesis position. Thus, it is a forecast of genesis at the wrong time.
Citation: Weather and Forecasting 28, 6; 10.1175/WAF-D-13-00008.1
False alarm (FA)—TC genesis is forecast in the model, but it does not occur within 24 h of a BT time and is not located within 5° of any BT location; these are cases that cannot be classified as a hit, EG, or LG.
The suggested maximum temporal tolerance for hit events that would be useful for operations is ± 24 h (R. Pasch 2012, personal communication), and a tolerance of ± 12 h would have reduced the POD and critical success index (CSI) (not shown). Thus, the ±24-h tolerance was selected. The EG and LG cases collectively are denoted as incorrect timing (IT) events. One also must consider the scenario when a TC in the BT dataset is not predicted by the models. Therefore, in addition to calculating our conditional probabilities given model-indicated genesis (hit, FA, EG, or LG), we also calculate metrics that account for missed events, such as the POD, bias, and CSI.
d. Statistical methods
Two-by-two contingency tables often are used when calculating skill scores for events that either do or do not occur (Fig. 5a). However, because we also include events having incorrect timing, the typical 2 × 2 contingency table was modified (Fig. 5b), where a is a hit, b is an FA, c is a miss, d is a correct null, e is an EG event, and f is a LG event. We adopt the following notation for our statistical calculations:
actual genesis occurs, Gy;
actual genesis does not occur, Gn;
actual genesis occurs but is forecast too early, Ge;
actual genesis occurs but is forecast too late, Gl;
actual genesis occurs but at the incorrect forecast time, Gt = Ge + Gl;
model-indicated genesis occurs, Fy; and
model-indicated genesis does not occur, Fn.
(a) Traditional 2 × 2 contingency table based on Wilks (2006). (b) Modified contingency table that accounts for incorrect timing events.
Citation: Weather and Forecasting 28, 6; 10.1175/WAF-D-13-00008.1
Our first set of statistics is conditioned on the forecast model predicting genesis. These conditional probabilities are defined as
P(Gy | Fy), the probability of actual genesis given that the model indicates genesis (hit);
P(Gn | Fy), the probability of no actual genesis given that the model indicates genesis (FA);
P(Ge | Fy), the probability of actual genesis but the model predicts genesis too early (EG); and
P(Gl | Fy), the probability of actual genesis but the model predicts genesis too late (LG).
We next consider cases when model-indicated genesis does not occur (Fn). Since the base state of the NATL is that a TC does not exist, correct null events would greatly outnumber the hit, FA, miss, and IT events, rendering the skill score calculations impractical. Therefore, we focus on missed events, P(Gy | Fn).
We analyze model performance in two ways: 1) the aforementioned conditional probabilities, which are the most useful information in a real-time setting, and 2) the traditional metrics on the performance diagram (POD, SR, bias, CSI), which provide an overall description of model performance and highlight where strengths and deficiencies exist. For reference, appendix B includes a list of the statistical terminology and corresponding acronyms and equations/symbols.
3. Results
We present results for each model separately, calculating the yearly conditional probability of each event type given that model-indicated genesis occurs (e.g., Fig. 6). This metric provides insight into how often TC genesis forecasts verify versus how often no TC develops. We then plot the events on a performance diagram (e.g., Fig. 7), which efficiently depicts several traditional success metrics in one figure. Each event also is plotted geographically (e.g., Fig. 8). We can use the geographic distribution of genesis events to determine a model's preferred genesis region(s). Finally, we calculate the conditional probability of each event type by forecast hour (e.g., Fig. 9) to show how the reliability of TC genesis forecasts evolves with time during the model cycle.
Conditional probability of a hit (solid line), FA (dashed line), EG (dotted line), and LG (dotted–dashed line) given model-indicated genesis from the CMC for each season, 2004–11. Genesis events from all forecast hours (6–96) are included. Numbers in parentheses are sample sizes.
Citation: Weather and Forecasting 28, 6; 10.1175/WAF-D-13-00008.1
Performance diagram for the CMC. SR is given along the x axis; POD on the y axis. Bias values are indicated by the dashed lines, and the CSI values are indicated by the curved, solid lines. A “perfect” performing model would be in the top-right corner of the plot. Genesis events from all forecast hours (6–96) are included.
Citation: Weather and Forecasting 28, 6; 10.1175/WAF-D-13-00008.1
The 2004–11 CMC hit (green triangle), FA (red square), and IT (blue circle) event locations. Numbers in parentheses are the numbers of model-indicated events. Outlined polygons represent subregions where conditional probabilities were calculated. These conditional probabilities are given in the bottom portion of the figure.
Citation: Weather and Forecasting 28, 6; 10.1175/WAF-D-13-00008.1
Conditional probabilities by forecast hour for the CMC. Lines and numbers below the forecast hours are as in Fig. 6.
Citation: Weather and Forecasting 28, 6; 10.1175/WAF-D-13-00008.1
a. CMC
Figure 6 shows the yearly conditional probability of a hit, FA, EG, and LG given that model-indicated genesis occurs. Since 2007, P(Gy | Fy) has increased, reaching its greatest seasonal value of 0.38 in 2011. FAs have been an issue, but improvements have occurred recently.
The performance diagram for the CMC (Fig. 7) reveals that the model predicted TC genesis rather aggressively during 2007 and 2008, likely due to a model upgrade in late 2006 when the horizontal grid spacing was changed from 100 to 33 km, the vertical resolution was increased from 28 to 58 levels, and changes to the convective schemes were made (Bélair et al. 2009). Bias values exceeding 1 and SR values less than 0.4 indicate that many of these genesis events were FAs. Improvements began in 2009, with SR values approaching 0.5 by 2011. The model also became less aggressive in 2011, as evidenced by the reduced bias (0.94 in 2010 and 0.57 in 2011). The CSI also generally has improved in recent years (below 0.13 during 2004–07; above 0.2 during 2008–11).
The locations of hits, FAs, and ITs are shown in Fig. 8. Unlike some of the other models, the CMC does not seem to have a distinct preferred region of genesis; genesis events occur over a larger portion of the basin compared to the other models. The model performs best over the main development region (MDR; 10°–20°N, 60°–20°W), Caribbean Sea, and Gulf of Mexico (GOM), where P(Gn | Fy) is smallest (~0.4–0.45). Poleward of 20°N and equatorward of 10°N, P(Gn | Fy) is much greater than P(Gy | Fy).
The locations of genesis events generally agree with the climatologically preferred areas of development. Most genesis events that develop off the coast of Africa occur during August and September, consistent with Gray (1968). However, during June, July, October, and November, genesis events generally occur closer to North America (not shown).
Figure 9 depicts the conditional probability of a hit, FA, EG, and LG by forecast hour. As expected, model performance decreases with increasing forecast hour. Approximately 94% of TC genesis events verify as hits at forecast hour 6, compared to only ~14% at forecast hour 96. Also note the sharp increase in the number of model-indicated genesis events between forecast hours 6 and 24. We speculate this may occur because the model may take several hours of integration before it is able to spin up a TC in most cases. The increased number of genesis events after forecast hour 24 is sustained throughout the remainder of the forecast period.
b. ECMWF
Figure 10 depicts results from the ECMWF model. Conditional probabilities and the number of total genesis events are quite variable from year to year, likely due to the small sample size of model-indicated genesis events. The range of P(Gy | Fy) is from 0.24 in 2007 to 0.52 in 2010. While there are fluctuations in P(Gy | Fy), the increase in P(Gt | Fy) during the past two seasons has contributed to a very low P(Gn | Fy) (e.g., 0.16 during 2011). The performance diagram in Fig. 11 shows that the ECMWF misses many genesis events (evidenced by the relatively low POD values), but when the model does predict genesis, development usually occurs.
As in Fig. 6, but for the ECMWF.
Citation: Weather and Forecasting 28, 6; 10.1175/WAF-D-13-00008.1
As in Fig. 7, but for the ECWMF.
Citation: Weather and Forecasting 28, 6; 10.1175/WAF-D-13-00008.1
The ECMWF exhibits preferred regions of genesis (Fig. 12). Over 60% of all events occur over the MDR. Relatively few events occur in other regions (e.g., GOM, east-central Atlantic). Although the ECMWF misses many TCs in the GOM (as do the other models), when it does forecast genesis there, it almost always occurs. In this region P(Gy | Fy) is 0.65, P(Gt | Fy) is 0.31, and P(Gn | Fy) is only 0.04. It also has a fairly low P(Gn | Fy) across the MDR and Caribbean Sea (~0.2). The model performs poorly over the east-central Atlantic. In general, the distribution of ECMWF TC genesis events follows the observed seasonal cycle. Early season (June–July) events occur mainly over the GOM and western Caribbean Sea. During the peak of the season (August/September) events occur throughout the basin, but mainly over the MDR. Toward the end of the season, events are confined to areas near the Antilles (not shown).
As in Fig. 8, but for the ECWMF.
Citation: Weather and Forecasting 28, 6; 10.1175/WAF-D-13-00008.1
The conditional probabilities by forecast hour are somewhat unusual (Fig. 13). Some models (e.g., CMC) show a steady increase in FA percentage and an increase in the total number of genesis events with increasing forecast hour. Conversely, the ECMWF shows the lowest probability of an FA during the first 18 forecast hours; beyond that, the performance oscillates, and the total number of genesis events fluctuates little with forecast hour.
As in Fig. 9, but for the ECWMF.
Citation: Weather and Forecasting 28, 6; 10.1175/WAF-D-13-00008.1
c. GFS
Figure 14 contains results for the GFS model. Performance noticeably improves during the 2010 season, possibly due to a major model upgrade during 2010 (appendix A). The horizontal resolution, planetary boundary layer scheme, and the deep and shallow convection schemes were just a few of the components upgraded (Environmental Modeling Center 2012). The results for 2011 are quite similar to those for 2010. During 2010 and 2011, P(Gy | Fy) ~0.48, P(Gt | Fy) ~0.14, and P(Gn | Fy) ~0.38. Thus, forecasters should have increased confidence in the GFS for predicting TC genesis.
As in Fig. 6, but for the GFS.
Citation: Weather and Forecasting 28, 6; 10.1175/WAF-D-13-00008.1
The improvement starting in 2010 also is clearly evident in the performance diagram (Fig. 15). SR nearly doubles and POD also increases. Bias scores indicate that GFS underpredicts genesis and that missed events are still an issue. But, when the GFS does forecast genesis, those forecasts more often verify as hits than FAs.
As in Fig. 7, but for the GFS.
Citation: Weather and Forecasting 28, 6; 10.1175/WAF-D-13-00008.1
The GFS exhibits preferred regions of TC genesis (Fig. 16). Nearly 60% of the genesis events occur over the MDR, like the ECMWF (Fig. 12). A particularly intriguing feature unique to the GFS is the number of FAs over the eastern NATL equatorward of 10°N. It appears that the GFS had a tendency to develop TCs out of the persistent areas of convection associated with the ITCZ. This issue appears to have been resolved by the 2010 model upgrade. During 2004–09, 48 FAs occur in this region, but during 2010–11 there are only 3 FAs in this area. Also noteworthy is the lack of FAs over the GOM, where P(Gn | Fy) is ~0.15. Model performance over the MDR, Caribbean Sea, and NW Atlantic are comparable [P(Gn | Fy) of 0.5–0.55]. Like the other models, the GFS performs poorly over the east-central Atlantic.
As in Fig. 8, but for the GFS.
Citation: Weather and Forecasting 28, 6; 10.1175/WAF-D-13-00008.1
Considering the entire 2004–11 period, the GFS exhibits a P(Gy | Fy) of 0.33 over the MDR, better than the basin-wide P(Gy | Fy) of 0.28. The MDR also is an area of marked improvement in TC genesis forecasts. In the MDR during 2010–11, the GFS exhibits a P(Gy | Fy) of 0.55, much better than the 8-yr average. Thus, a potential contributing factor to the GFS's improvement in 2010 and 2011 is that several BT TCs formed over the MDR, one of the model's preferred regions for TC genesis.
The GFS approximately follows the seasonal cycle of preferred genesis locations outlined in Gray (1968). Most of the genesis events off the coast of Africa occur during August and September, while most of the genesis events over the western Caribbean Sea and near the Bahamas occur either earlier or later in the year (not shown).
Like the CMC, the GFS also exhibits more genesis events at the later forecast hours. The increase occurs at ~36–48 h, then the number of TC genesis events plateaus for the remainder of the forecast period (Fig. 17). The value of P(Gy | Fy) decreases with increasing forecast hour, as one would expect. However, P(Gn | Fy) from forecast hours 30 to 78 generally remains between 0.5 and 0.6. This implies that the GFS does not predict a greater percentage of FAs during this period, but that the percentage of IT events increases with increasing forecast hour at the expense of the hit percentage. Beyond forecast hour 78, GFS performance steadily declines.
As in Fig. 9, but for the GFS.
Citation: Weather and Forecasting 28, 6; 10.1175/WAF-D-13-00008.1
d. NOGAPS
Figure 18 contains conditional probabilities for the NOGAPS model. During 2004–07, P(Gn | Fy) is quite high but then decreases, while P(Gy | Fy) peaks in 2009. NOGAPS does not successfully predict any genesis events during 2004–06 despite the major TC activity during the 2004–05 seasons. While the CMC produces a large number of FAs and sometimes overpredicts genesis (Fig. 7), NOGAPS exhibits the opposite problem (Fig. 19). The CSI has improved in recent years, but still is smaller than the CSI for the other models, both during 2011 and averaged over the entire period of study.
As in Fig. 6, but for the NOGAPS.
Citation: Weather and Forecasting 28, 6; 10.1175/WAF-D-13-00008.1
As in Fig. 7, but for the NOGAPS.
Citation: Weather and Forecasting 28, 6; 10.1175/WAF-D-13-00008.1
NOGAPS does not exhibit a distinct preferred region of genesis (Fig. 20), unlike the ECMWF and GFS models, which clearly favor the MDR. Most of the NOGAPS genesis events occur over the MDR, Caribbean Sea, and NW Atlantic, but very few occur over the east-central Atlantic and equatorward of 10°N.
As in Fig. 8, but for the NOGAPS.
Citation: Weather and Forecasting 28, 6; 10.1175/WAF-D-13-00008.1
NOGAPS exhibits seasonal preferences for TC genesis. During June and July there are very few genesis events. The majority of the genesis events occur during August and September, coincident with the climatological peak of TC activity. These events occur throughout the basin. During October and November most genesis events occur in the western Caribbean Sea and GOM, consistent with the climatologically favored regions of development for those months (not shown).
The small sample size of NOGAPS genesis events as a function of forecast hour does not support any definitive conclusions (Fig. 21). However, the model appears more likely to develop a TC later in the forecast period, and P(Gn | Fy) increases with increasing forecast hour, much like the CMC and GFS.
As in Fig. 9, but for the NOGAPS.
Citation: Weather and Forecasting 28, 6; 10.1175/WAF-D-13-00008.1
e. UKMET
Figure 22 provides verification statistics for the UKMET model. The model performed anomalously well in 2006 and anomalously poorly in 2007. Otherwise, it has shown relatively steady improvement, with P(Gy | Fy) increasing and P(Gn | Fy) decreasing. The performance diagram (Fig. 23) confirms that the improvements have occurred recently (i.e., 2009–11 are the years with the highest SR and CSI).
As in Fig. 6, but for the UKMET.
Citation: Weather and Forecasting 28, 6; 10.1175/WAF-D-13-00008.1
As in Fig. 7, but for the UKMET.
Citation: Weather and Forecasting 28, 6; 10.1175/WAF-D-13-00008.1
The geographic distribution of events is shown in Fig. 24. Over two-thirds of all model-indicated genesis events occur over the MDR. Over the MDR, P(Gn | Fy) is 0.26, behind only the ECMWF (0.21). Like the GFS and ECMWF, the UKMET misses many events over the GOM, but it also has a low P(Gn | Fy) there (0.21). UKMET performs poorly over the east-central Atlantic, as do the other models. The UKMET seems to follow the seasonal cycle of favorable TC genesis locations described by Gray (1968). Very few genesis events occur during June and July, while the majority of genesis events occur during August and September over the MDR. Some events occur during October and November, mainly over the Caribbean Sea and GOM (not shown).
As in Fig. 8, but for the UKMET.
Citation: Weather and Forecasting 28, 6; 10.1175/WAF-D-13-00008.1
Figure 25 shows the results by forecast hour. The UKMET performs best up to forecast hour 24, with the TC genesis forecasts degrading at later times. There is a spike in the number of genesis events at forecast hour 84, but otherwise, the number of events as a function of forecast hour exhibits no discernible trend, unlike the CMC and GFS.
As in Fig. 9, but for the UKMET.
Citation: Weather and Forecasting 28, 6; 10.1175/WAF-D-13-00008.1
f. Consensus TC genesis forecasting
Operational forecasters examine a suite of models when forecasting TC track and intensity. When the models are in good agreement, the forecaster has increased confidence in the forecast. Similarly, forecasters are likely to have increased confidence that a TC will develop when forecast by more than one model. We investigated whether better genesis forecasts occur when predicted by multiple models.
We documented all cases when two or more models at the same initialization time predicted genesis within 5° of each other (considered the same genesis event). We then determined whether the genesis forecast was a hit, FA, or IT event. Based on the forecast location and time of genesis, the genesis event could be a hit in one model, but a FA or IT in another model. Several experiments were conducted with different combinations of models (Table 5).
Multimodel consensus forecasts showing P(Gy | Fy) for all models included (i.e., all model forecasts verify as a hit) and P(Gn | Fy) for all models included (i.e., genesis does not occur). Sample size indicates the number of model runs.
Since each model has different strengths and weaknesses, they did not always agree on a given genesis event. For example, since the NOGAPS model is especially prone to underpredict genesis, there were fewer cases when it was part of a multiple model genesis forecast compared to a more aggressive model like the CMC. In fact, the CMC and UKMET provided the greatest multimodel dataset to examine (121 cases). Both forecasts verified as hits 31% of the time, and both forecasts verified as FAs 16% of the time. In the remaining 53% of cases, the forecast was a hit or IT in one model, and a FA or IT in the other model. Thus, if the CMC and UKMET both predict the same genesis event, there is a 16% chance that nothing will develop (i.e., both FAs). Conversely, there is an 84% chance that genesis will occur, although the models may incorrectly forecast the genesis time. Of all two-model combinations, the GFS–UKMET consensus yielded the highest probability of a hit for both models (48%). But, the ECMWF–NOGAPS consensus yielded the lowest probability of a FA for both models (6%).
We also examined multiple model forecasts of genesis with three-model combinations. As expected, results from the three-model combinations generally are better than from the two-model combinations. Among the best are the GFS–NOGAPS–UKMET and the ECMWF–GFS–NOGAPS combinations, which yield a 100% chance that genesis does occur (albeit potentially at the wrong time). This result is similar to that found by Vislocky and Fritsch (1995) when examining consensus forecasts for model output statistics (MOS). They determined that MOS from a model that may have a relatively poor level of performance by itself is still useful when part of a multimodel consensus. Unfortunately, it is uncommon for three models to predict the same genesis event. Rarer still are the four-model combinations. All of the four-model combinations yield a 0% chance that nothing will develop, except for the CMC–ECMWF–NOGAPS–UKMET combination (7%). There are only 10 instances when all five models predict the same genesis event. As expected, genesis ultimately occurs in each case.
Overall, if two or more models predict TC genesis, there is a relatively low probability that all of the models are predicting an FA. Thus, the results from the various consensus experiments are better than those from any single model. A forecaster should have increased confidence in a TC genesis forecast when predicted by multiple models—even if one of those models typically is aggressive or performs relatively poorly on its own.
4. Summary and conclusions
Forecasts of TC genesis from five global models (CMC, ECMWF, GFS, NOGAPS, and UKMET) were verified over eight seasons in the NATL basin (except five seasons for ECMWF). Each TC genesis event was identified, tracked, and classified as a hit, FA, EG, or LG event. We also considered missed events.
The results revealed several commonalities among the models. The conditional probability of a hit, given model-indicated genesis, was greatest during 2011 for the CMC and GFS. However, we cannot determine whether this result is due to improvements in the models, environmental conditions (e.g., actual storms developing in a region where the models are more likely to predict genesis, such as the MDR), or simply coincidence. Analyzing future seasons will help answer that question. Each model appeared to generally follow the observed climatological cycle of the TC season. That is, relatively few genesis events occurred during June, July, October, and November. Those that did occur typically were located over the Caribbean Sea, GOM, and near the southeast United States. The majority of TC genesis events occurred during August and September. Very few genesis events were forecast over the GOM, indicating that the models typically miss genesis in that area. Given the disparity in the number of genesis events between the MDR and GOM, it is possible that the models generally have more difficulty predicting the tropical transition and eventual genesis of a TC with baroclinic origins (typically found in the GOM) compared to TC genesis from an AEW. There also were few FAs over the eastern Caribbean Sea, consistent with the climatological scarcity of genesis in that region (Shieh and Colucci 2010). All of the models performed poorly in the east-central Atlantic region. The conditional probability of a FA generally increased with increasing forecast hour in most of the models, confirming the perception that TC genesis forecasts are less reliable at longer forecast lead times.
The models also exhibited clear differences. The CMC was the most aggressive in forecasting TC genesis. As a result, it had the greatest number of hits, but also the greatest number of FAs. NOGAPS was least aggressive, underpredicting TC genesis during all seasons. Some of the models had a clear preferred region for genesis (ECMWF, GFS, and UKMET), while others showed genesis events spread more evenly throughout the basin (CMC and NOGAPS).
Several metrics were used to rank the models, with a comparison given in Fig. 26. The performance diagram shows that the 2010–11 average performance is better than the 2007–11 average for all models, indicating that the models are improving over time. In terms of the conditional probability of a hit given model-indicated genesis [P(Gy | Fy)] averaged over 2007–11 (the years when output from all models were available), the ECMWF was best (0.44), followed by the UKMET (0.36), the GFS (0.34), the CMC (0.27), and the NOGAPS (0.26). However, in terms of CSI, the CMC was best (0.21), followed by the GFS and the UKMET (0.14), the ECMWF (0.1), and the NOGAPS (0.06). Thus, when considering long-term averages, the relative performance depends on which metric one chooses. The NOGAPS performed relatively poorly individually, mainly due to its extreme underprediction of TC genesis. However, its conservative nature is useful when part of a multimodel consensus forecast.
Performance diagram with the average 2007–11 and 2010–11 performances for each model. Genesis events from all forecast hours (6–96) are included.
Citation: Weather and Forecasting 28, 6; 10.1175/WAF-D-13-00008.1
Experiments with consensus forecasts revealed that FA rates were lower when multiple models predicted TC genesis. The consensus approach yielded better results than any one model alone. For example, when the ECMWF and NOGAPS both predict the same genesis event, there is a 94% chance that it will occur (albeit potentially at the wrong time). This argues that forecasters should have greater confidence in TC genesis forecasts made by multiple models. Unfortunately, the sample size of such consensus forecasts is relatively small.
The top-performing model during one season may not be the top performer the next season (Fig. 27). An upgrade to a given model may improve its performance such that it ranks better than its counterparts during the next season. Also, the locations of actual genesis cases may play a role in which model performs best. For example, a season where most TCs form from AEWs over the MDR may bolster the performance of the ECMWF, GFS, or UKMET, where the MDR is a preferred region of genesis. If most TCs develop in the Caribbean Sea or GOM, that may favor the CMC, which has a less distinct regional preference for genesis. Therefore, it is impossible to predict which model will be the most reliable during an upcoming season. Nonetheless, our approach does indicate the models' past performance and whether they have been improving or degrading.
The conditional probability of a hit [P(Gy | Fy)] for each model by season. Genesis events from all forecast hours (6–96) are included.
Citation: Weather and Forecasting 28, 6; 10.1175/WAF-D-13-00008.1
In the future, we plan to expand this analysis to other basins and to investigate why the models perform as they do. We will attempt to determine which changes or upgrades in the model yielded changes in model performance. We also will seek to determine whether particular synoptic situations are associated with relatively good or poor model performance. The ultimate goal of this research is to develop statistically based genesis probabilities based on additional predictors to aid operational forecasters in predicting TC genesis.
Acknowledgments
The authors thank Ryan Truchelut, who provided some of the initial ideas for this research. We are also indebted to Julian Heming of the Met Office for providing some of the model output and upgrade information that was necessary to carry out this research. ECMWF output was obtained from the TIGGE archive (ECMWF portal). We thank the anonymous reviewers, as well as Chris Landsea and Dave Zelinsky, who conducted the NHC internal review. Thanks to Ron McTaggart-Cowan and Dave Ryglicki for providing references for model upgrades. The authors also benefited from discussions with Pat Harr at NPS and Mike Fiorino at ESRL regarding this research topic and from the suggestion by Sim Aberson at HRD to display some of our results using the performance diagrams. Finally, this research was supported by NOAA/COMET Partner's Project Z12-93225, NASA GRIP Grant NNX09AC43G, and a Florida State University fellowship.
APPENDIX A
Selected Model Upgrades
a. CMC (Bélair et al. 2009; Charron et al. 2012)
Implemented 31 October 2006
Horizontal grid spacing reduced from 100 to 33 km
Number of vertical levels increased from 28 to 58
Time step reduced from 45 to 15 min
Kuo transient shallow convection scheme added (none previously)
Deep convection scheme changed from Kuo to Kain–Fritsch
Implemented 22 June 2009
Vertical coordinate changed from normalized σ to hybrid σ pressure
Number of vertical levels increased from 58 to 80
Model top raised from 10 to 0.1 hPa
b. ECMWF (2012)
June 2007—replaced shortwave radiation scheme with Rapid Radiative Transfer Model
November 2007—updated convection scheme
November 2007—updated vertical diffusion above the boundary layer
June 2008—updated moist physics in the tangent linear and adjoint models used in the four-dimensional variational data assimilation (4DVAR) scheme
September 2008—used high-resolution SST from the Met Office
September 2008—introduced a simple representation of the diurnal cycle of SST
September 2009—introduced a new approach for quality controlling conventional observations
January 2010—horizontal resolution changed from T799 (~25 km) to T1279 (~16 km)
November 2010—new cloud scheme implemented
c. GFS (Environmental Modeling Center 2012)
May 2005—horizontal resolution changed from T254L64 to 84 h and T170L42 to 180 h to T382L64 to 180 h
May 2007—changed to a hybrid σ-pressure vertical coordinate
May 2007—unified the NCEP 3DVAR system under the gridpoint statistical interpolation (GSI)
July 2010—horizontal resolution changed to T574L64 to 192 h
July 2010—made changes to the hurricane relocation algorithm
July 2010—upgraded boundary layer scheme
July 2010—new mass flux shallow convection scheme added
July 2010—updated deep convection scheme
d. NOGAPS (NRL 2012; COMET 2012)
September 2004—changes made to the convective momentum transport within the Emanuel parameterization
2005—TC bogus scheme improved
September 2009—model resolution increased from T239L30 to T239L42
September 2009—switched from a 3DVAR data assimilation system to the 4DVAR extension of the Naval Research Laboratory Atmospheric Variational Data Assimilation System (NAVDAS-AR)
September 2010—model resolution increased from T239L42 to T319L42
2010—vertical hybrid coordinate implemented
e. UKMET (J. Heming 2012, personal communication)
October 2004—introduction of 4DVAR
January 2005—model physics upgraded
December 2005—horizontal grid spacing decreased to 40 km
December 2005—increase to 50 vertical levels
December 2006—GPS radio occultation (GPSRO) data introduced
May 2007—model physics upgraded
November 2008—model physics upgraded
November 2009—model physics and dynamics revised
November 2009—increase to 70 vertical levels
March 2010—horizontal grid spacing decreased to 25 km
July 2010—new cloud scheme introduced
July 2011—hybrid data assimilation introduced
APPENDIX B
Terms referenced in the text and their associated acronyms and equations/symbols.
REFERENCES
Bélair, S., Roch M. , Leduc A.-M. , Vaillancourt P. A. , Laroche S. , and Mailhot J. , 2009: Medium-range quantitative precipitation forecasts from Canada's new 33-km deterministic global operational system. Wea. Forecasting, 24, 690–708.
Beven, J. L., 1999: The Boguscane—A serious problem with the NCEP Medium Range Forecast model in the tropics. Preprints, 23rd Conf. on Hurricanes and Tropical Meteorology, Dallas, TX, Amer. Meteor. Soc., 845–848.
Braun, S. A., and Coauthors, 2013: NASA's Genesis and Rapid Intensification Processes (GRIP) field experiment. Bull. Amer. Meteor. Soc., 94, 345–363.
Briegel, L. M., and Frank W. M. , 1997: Large-scale influences on tropical cyclogenesis in the western North Pacific. Mon. Wea. Rev., 125, 1397–1413.
Camargo, S. J., Emanuel K. A. , and Sobel A. H. , 2007: Use of a genesis potential index to diagnose ENSO effects on tropical cyclone genesis. J. Climate, 20, 4819–4834.
Chan, J. C. L., and Kwok R. H. F. , 1999: Tropical cyclone genesis in a global numerical weather prediction model. Mon. Wea. Rev., 127, 611–624.
Charney, J. G., and Eliassen A. , 1964: On the growth of the hurricane depression. J. Atmos. Sci., 21, 68–75.
Charron, M., and Coauthors, 2012: The stratospheric extension of the Canadian Global Deterministic Medium-Range Weather Forecasting System and its impact on tropospheric forecasts. Mon. Wea. Rev., 140, 1924–1944.
Cheung, K. W., and Elsberry R. L. , 2002: Tropical cyclone formations over the western North Pacific in the Navy Operational Global Atmospheric Prediction System Forecasts. Wea. Forecasting, 17, 800–820.
COMET, cited 2012: NOGAPS 4.0 introduction. UCAR COMET Program. [Available online at http://www.meted.ucar.edu/nwp/pcu2/nogaps/index.htm.]
Côté, J., Desmarais J. G. , Gravel S. , Méthot A. , Patoine A. , Roch M. , and Staniforth A. , 1998a: The operational CMC–MRB Global Environmental Multiscale (GEM) model. Part II: Results. Mon. Wea. Rev., 126, 1397–1418.
Côté, J., Gravel S. , Méthot A. , Patoine A. , Roch M. , and Staniforth A. , 1998b: The operational CMC–MRB Global Environmental Multiscale (GEM) model. Part I: Design considerations and formulation. Mon. Wea. Rev., 126, 1373–1395.
Craig, G. C., and Gray S. L. , 1996: CISK or WISHE as the mechanism for tropical cyclone intensification. J. Atmos. Sci., 53, 3528–3540.
Cullen, M. J. P., 1993: The Unified Forecast/Climate Model. Meteor. Mag., 122, 81–94.
Dunkerton, T. J., Montgomery M. T. , and Wang Z. , 2008: Tropical cyclogenesis in a tropical wave critical layer: Easterly waves. Atmos. Chem. Phys., 8, 11 149–11 292.
ECMWF, cited 2012: ECMWF annual reports. [Available online at http://www.ecmwf.int/publications/annual_report/.]
Elsberry, R. L., Clune W. M. , and Harr P. A. , 2009: Evaluation of global model early track and formation prediction in the western North Pacific. Asia-Pac. J. Atmos. Sci., 45, 357–374.
Elsberry, R. L., Jordan M. S. , and Vitart F. , 2010: Predictability of tropical cyclone events on intraseasonal timescales with the ECMWF monthly forecast model. Asia-Pac. J. Atmos. Sci., 46, 135–153.
Elsberry, R. L., Jordan M. S. , and Vitart F. , 2011: Evaluation of the ECMWF 32-day ensemble predictions during 2009 season of western North Pacific tropical cyclone events on intraseasonal timescales. Asia-Pac. J. Atmos. Sci., 47, 305–318.
Emanuel, K. A., 1986: An air–sea interaction theory for tropical cyclones. Part I: Steady-state maintenance. J. Atmos. Sci., 43, 585–605.
Emanuel, K. A., 1988: The maximum intensity of hurricanes. J. Atmos. Sci., 45, 1143–1155.
Emanuel, K. A., and Nolan D. S. , 2004: Tropical cyclones and the global climate system. Preprints, 26th Conf. on Hurricanes and Tropical Meteorology, Miami Beach, FL, Amer. Meteor. Soc., 10A.1. [Available online at https://ams.confex.com/ams/pdfpapers/75463.pdf.]
Enagonio, J., and Montgomery M. T. , 2001: Tropical cyclogenesis via convectively forced vortex Rossby waves in a shallow water primitive equation model. J. Atmos. Sci., 58, 685–706.
Environmental Modeling Center, cited 2012: GFS/GDAS changes since 1991. [Available online at http://www.emc.ncep.noaa.gov/gmb/STATS/html/model_changes.html.]
Frank, W. M., and Young G. S. , 2007: The interannual variability of tropical cyclones. Mon. Wea. Rev., 135, 3587–3598.
Goerss, J. S., 2000: Tropical cyclone track forecasts using an ensemble of dynamical models. Mon. Wea. Rev., 128, 1187–1193.
Goerss, J. S., Sampson C. R. , and Gross J. M. , 2004: A history of western North Pacific tropical cyclone track forecast skill. Wea. Forecasting, 19, 633–638.
Gray, W. M., 1968: Global view of the origin of tropical disturbances and storms. Mon. Wea. Rev., 96, 669–700.
Jarvinen, B. R., Neumann C. J. , and Davis M. A. S. , 1984: A tropical cyclone data tape for the North Atlantic basin, 1886–1983: Contents, limitations, and uses. NWS NHC Tech. Memo. 22, 24 pp.
Kanamitsu, M., 1989: Description of the NMC global data assimilation and forecast system. Wea. Forecasting, 4, 335–342.
Marchok, T. P., 2002: How the NCEP tropical cyclone tracker works. Preprints, 25th Conf. on Hurricanes and Tropical Meteorology, San Diego, CA, Amer. Meteo. Soc., P1.13. [Available online at https://ams.confex.com/ams/25HURR/webprogram/Paper37628.html.]
McAdie, C. J., Landsea C. W. , Neumann C. J. , David J. E. , Blake E. , and Hammer G. R. , 2009: Tropical cyclones of the North Atlantic Ocean, 1851–2006. Historical Climatology Series 6-2, National Climatic Data Center–NHC, 238 pp.
McTaggart-Cowan, R., Deane G. D. , Bosart L. F. , Davis C. A. , and Galarneau T. J. , 2008: Climatology of tropical cyclogenesis in the North Atlantic (1948–2004). Mon. Wea. Rev., 136, 1284–1304.
Montgomery, M. T., and Enagonio J. , 1998: Tropical cyclogenesis via convectively forced vortex Rossby waves in a three-dimensional quasigeostrophic model. J. Atmos. Sci., 55, 3176–3207.
Montgomery, M. T., Nguyen V. S. , Persing J. , and Smith R. K. , 2009: Do tropical cyclones intensify by WISHE? Quart. J. Roy. Meteor. Soc., 135, 1697–1714.
Montgomery, M. T., and Coauthors, 2012: The Pre-Depression Investigation of Cloud-Systems in the Tropics (PREDICT) experiment: Scientific basis, new analysis tools, and some first results. Bull. Amer. Meteor. Soc., 93, 153–172.
NHC, cited 2011: NHC track and intensity models. [Available online at http://www.nhc.noaa.gov/modelsummary.shtml.]
NRL, cited 2012: Marine Meteorology Division history. [Available online at http://www.nrlmry.navy.mil/MMD_History/text/frame.htm.]
Pasch, R. J., Harr P. A. , Avila L. A. , Jiing J. G. , and Elliot G. , 2006: An evaluation and comparison of predictions of tropical cyclogenesis by three global forecast models. Preprints, 27th Conf. on Hurricanes and Tropical Meteorology, Monterey, CA, Amer. Meteor. Soc., 14B.5. [Available online at https://ams.confex.com/ams/pdfpapers/108725.pdf.]
Pasch, R. J., Blake E. S. , Jiing J. G. , Mainelli M. M. , and Roberts D. P. , 2008: Performance of the GFS in predicting tropical cyclone genesis during 2007. Preprints, 28th Conf. on Hurricanes and Tropical Meteorology, Orlando, FL, Amer. Meteor. Soc., 11A.7. [Available online at https://ams.confex.com/ams/28Hurricanes/techprogram/paper_138218.htm.]
Rappaport, E. N., and Coauthors, 2009: Advances and challenges at the National Hurricane Center. Wea. Forecasting, 24, 395–419.
Ritchie, E. A., and Holland G. J. , 1997: Scale interactions during the formation of Typhoon Irving. Mon. Wea. Rev., 125, 1377–1396.
Roebber, P. J., 2009: Visualizing multiple measures of forecast quality. Wea. Forecasting, 24, 601–608.
Rogers, R., and Coauthors, 2006: The Intensity Forecasting Experiment: A NOAA multiyear field program for improving tropical cyclone intensity forecasts. Bull. Amer. Meteor. Soc., 87, 1523–1537.
Rosmond, T. E., 1992: The design and testing of the Navy Operational Global Atmospheric Prediction System. Wea. Forecasting, 7, 262–272.
Sampson, C. R., and Schrader A. J. , 2000: The Automated Tropical Cyclone Forecasting System (version 3.2). Bull. Amer. Meteor. Soc., 81, 1231–1240.
Sampson, C. R., Franklin J. L. , Knaff J. A. , and DeMaria M. , 2008: Experiments with a simple tropical cyclone intensity consensus. Wea. Forecasting, 23, 304–312.
Shieh, O. H., and Colucci S. J. , 2010: Local minimum of tropical cyclogenesis in the eastern Caribbean. Bull. Amer. Meteor. Soc., 91, 185–196.
Simpson, J., Ritchie E. , Holland G. J. , Halverson J. , and Stewart S. , 1997: Mesoscale interactions in tropical cyclone genesis. Mon. Wea. Rev., 125, 2643–2661.
Tory, K. J., and Frank W. M. , 2010: Tropical cyclone formation. Global Perspectives on Tropical Cyclones: From Science to Mitigation, J. C. L. Chan and J. D. Kepert, Eds., World Scientific, 55–91.
Tsai, H.-C., Lu K.-C. , Elsberry R. L. , Lu M.-M. , and Sui C.-H. , 2011: Tropical cyclone–like vortices detection in the NCEP 16-day ensemble system over the western North Pacific in 2008: Application and forecast evaluation. Wea. Forecasting, 26, 77–93.
Vislocky, R. L., and Fritsch J. M. , 1995: Improved model output statistics forecasts through model consensus. Bull. Amer. Meteor. Soc., 76, 1157–1164.
Walsh, K. J. E., Fiorino M. , Landsea C. W. , and McInnes K. L. , 2007: Objectively determined resolution-dependent threshold criteria for the detection of tropical cyclones in climate models and reanalyses. J. Climate, 20, 2307–2314.
Wilks, D. S., 2006: Statistical Methods in the Atmospheric Sciences. 2nd ed. Academic Press, 627 pp.
ATCF system is a software package used by NHC and other operational centers to improve operational efficiency. The ATCF system provides TC “fix” observations and updates to the “working” best track, among other tasks (Rappaport et al. 2009). CARQ entries in the A-deck dataset are the hurricane specialists' operational initial best estimates of the location, intensity, and other parameters (e.g., initial motion) of a tropical disturbance or cyclone that are used as input to forecast models.