a. Previous work

Prior to Gray (1984b), extended-range forecasting of TC activity was limited and focused on issues related to potential predictability rather than actual attempted yearly forecasts. Ballenzweig (1959) distinguished active versus inactive TC months by studying composites of large-scale atmospheric fields. His report compiled and differenced active from inactive months during August–October and attempted to link the variable activity during these periods to circulation anomalies across the Northern Hemisphere (NH). He found that months of maximum TC activity were associated with a northeastward shift of the Atlantic anticyclone, thereby expanding and weakening the area of easterlies equatorward of the ridge. Whereas Ballenzweig's study was diagnostic rather than prognostic, it introduced and illustrated the notion of observable mean differences between active and inactive periods occurring on month-long time scales.

Another study dealing with short-term differences of TC activity was performed by Shapiro (1987) who tested the predictability of monthly Atlantic TC frequency using monthly mean wind and sea level pressure anomaly (SLPA) data poleward of 20°N. These parameters were calculated over the Atlantic basin 2 months in advance of the August–October active part of the hurricane season and examined for predictive associations. Shapiro found statistically significant correlations in TC frequency encompassing approximately 45% of the hindcast variance. His results suggested that the phase of the El Niño–Southern Oscillation (ENSO) was the strongest modulator of monthly Atlantic TC activity though he estimated that only one-sixth of his skill was directly related to ENSO.

Recent findings by Maloney and Hartmann (2000) indicate that hurricane activity in the Gulf of Mexico (GOM) and western Caribbean Sea is significantly modulated by the passage of the Madden–Julian oscillation (MJO). The MJO propagates eastward across the Tropics as a wavenumber 1 oscillation with a period of about 30–45 days. The authors noted a fourfold increase of western Caribbean and GOM genesis events when MJO-linked 850-mb u-wind anomalies were westerly across the eastern Pacific south of Mexico. Maloney and Hartmann state that more accurate predictions of week-to-week genesis events may be possible with forecasts of the MJO.

DeMaria et al. (2001) were likely the first to utilize daily data for real-time forecasting of intraseasonal variations in Atlantic TC activity. They created a “genesis parameter” derived from the 5-day mean of vertical wind shear, midlevel moisture, and vertical stability for the tropical Atlantic east of the Lesser Antilles. This genesis parameter explained about 50% of the variance of TC activity that formed in this area during 1995–99. The authors suggest that monitoring the genesis parameter on a daily basis provides an improved measure of the probability of a tropical wave developing into a TC.

b. Objective and outline

A forecast of August-only Atlantic basin TC activity is the focus of the remainder of this paper. A test of Gray and colleagues' seasonal forecast scheme explained only about 20% of the August NTC hindcast variance, and therefore new forecast techniques needed to be developed for shorter-period predictions. Tropical cyclone data and August climatology are the subjects of section 2, and the forecast methodology along with the global reanalysis data used are detailed in section 3. Section 4 details the physical relationships between predictors and TC activity, and section 5 examines hindcasting results of the August-only forecast scheme. Applications of the August-only forecast including a forecast of U.S. landfalling TCs are the topic of section 6. Section 7 details the conclusions and ideas for future work.

a. Tropical cyclone data

August TC information was calculated from the “best track” database maintained by the National Hurricane Center in Miami, Florida. All available track and intensity estimates were used to prepare a smoothed postanalysis composed of the most accurate storm information (Jarvinen et al. 1984; Neumann et al. 1999). Tropical cyclones whose life cycle spanned any portion of August were tallied and stratified by their maximum intensity during August. The basic activity parameters calculated were the number of days during August that a TC existed as a named storm (NS), hurricane (H), and intense hurricane (IH; Saffir–Simpson category 3 or greater). This method thus takes into account TCs that formed in July and persisted into August but considered only the portion of activity that occurred during August.

August TCs were also classified on the basis of their origins as suggested by Hess et al. (1995), which showed improvements in seasonal forecasting skill obtained by separately forecasting TC formations based on origin-linked classes. One such origin-linked class is termed tropical only (TO). The TO systems are defined as those hurricanes that formed without any obvious midlatitude influences, typically from African easterly waves equatorward of 23.5°N. An alternative class was termed baroclinically initiated (BI) hurricanes wherein nontropical disturbances were involved. The latter class may include storms that develop along a stationary front, from a decaying midlatitude low, or from a mesoscale convective system (MCS) emerging from North America (Elsner and Kara 1999). Categorizing TCs at the time they first reached tropical storm intensity as either TO or BI posed an additional problem since Hess et al. (1995) considered only hurricanes. It was decided that the key criterion for classifying an NS as a TONS was that it developed from a tropical wave without any contribution from midlatitude influences. Note however that a TONS that subsequently became a hurricane due to nontropical enhancement effects remained a TONS but was not considered a TOH. Regression techniques for both TO and BI formations were tested in this study though no BI forecasting technique proved successful.

b. August climatology

This section presents a summary of the main climatological qualities of August TC variability. August is the secondmost active month on average, encompassing about 26% of the total seasonal NTC. The latter value has ranged from a high of 67.6% (1983) to a low of 0% (1961, 1997) as shown in Table 1. On average, August has approximately three TCs, two of which become hurricanes (H), and one of which becomes an intense hurricane. Table 2 presents a summary of statistics for August activity. Note that 1955 and 1995 were the two most active Augusts while 1961 and 1997 had no August TC activity.

Table 3 contains the linear cross correlations between all August TC parameters. Though most of the indices in the table are closely associated, the NS and IH values are not well correlated. A large number of NSs in August is only weakly indicative that an above average number of IHs will form. An example is August 1995 when eight NS occurred, but only one of these reached IH status. This dichotomy suggests that conditions favorable for TC formation are not necessarily the same as conditions conducive for strengthening into powerful storms. It is not surprising that different sets of predictors are chosen to forecast the various TC indices, as detailed in section 5.

c. Multidecadal variations

Gray (1990) and Gray et al. (1997b) proposed that the recent 25-yr (1970–94) downturn of TC activity, especially IHs, was primarily due to variations in the broadscale Atlantic thermohaline circulation for which North Atlantic (50°–60°N, 10°–50°W) SSTs are used as a proxy. Multidecadal variations of this circulation and associated ocean SST patterns tend to occur on periods of 30 yr or more. Broadscale North Atlantic SSTs were in a relatively “cool” phase from the early 1970s to the early 1990s, which closely coincides with observed reduced TC activity. However, North Atlantic SSTs have warmed dramatically since mid-1995 with a return to an SST configuration and level of TC activity more closely resembling that of the 1950s and 1960s.

Multidecadal signals in the August monthly data are considered by compositing aggregate activity during the 1950–69 and 1995–99 time periods (both of these periods judged to be during an active Atlantic Ocean thermohaline circulation) versus the inactive 1970–94 period. Table 4 shows the difference between the two 25-yr periods expressed both as comparative averages and as ratios. The most pronounced changes are for the most intense type of activity, which was much greater in the active thermohaline period. Whereas August TO activity was notably suppressed during the inactive thermohaline years, the incidence of higher-latitude BI systems increased. Thirteen August BI hurricanes formed during the inactive era, while only one developed during the active era.

The development of the August-only TC climatology allowed the formulation of indices describing measures of August TC activity. Some of these August indices proved useful as precursor signs for the remainder of the season's activity.

a. Global data

Analyses performed in this study were based on the monthly National Centers for Environmental Prediction–National Center for Atmospheric Research (NCEP–NCAR) global reanalysis (Kalnay et al. 1996) data for 1949–99. The reanalysis is composed of numerous atmospheric and SST fields integrated onto a 2.5° × 2.5° global grid. The analysis technique gives a realistic, consistent interpretation of large-scale weather features back to 1948 using standardized assimilation and interpolation methods. This unique dataset allows for comparison of recent climate data versus data from earlier years in an environment free of artificial “climate shifts” due to inaccurate data or changes in most observational platforms or analysis procedures (Kalnay et al. 1996). However this does not take into account changes in observational platforms such as geostationary satellite networks affecting the monitoring of TCs.

The online data analysis and plotting resources at the Climate Diagnostics Center (CDC) facilitated the identification and evaluation of correlations between atmospheric data fields and August TC activity parameters. All TC statistics were tested for antecedent correlation with reanalysis data fields for prior months on the CDC Web page [http://www.cdc.noaa.gov/correlation; see Smith and Brown (1999)]. An example given in Fig. 3 shows the correlation field for July SLP associated with 1958–99 August NTC. The areas with large correlations in the North Pacific and North Atlantic Oceans in Fig. 3 were used as predictors after undergoing additional testing as detailed in the following section.

b. Methodology

It was clear that a new set of predictive factors would be required for skillful August-only forecasts as seasonal forecast factors failed to capture variability on monthly time scales. These new predictors were identified in several ways. First, the 10 most active Augusts in the data record were composited and differenced from the 10 least active Augusts. The CDC reanalysis compositing site was similarly helpful for delineating broad areas with noticeable circulation differences for active versus inactive Augusts on monthly time scales. As shown in the flow chart in Fig. 4, the methodology identified various global-scale difference fields, which were then tested further as potential predictors in a statistical sense, and later in a physical one. Another group of provisional predictors was identified by correlating TC activity indices with global reanalysis atmospheric fields for 1958–99. Additional predictors were found later by correlating forecast residuals with global reanalysis data.

Predictors thus identified were extracted from the full reanalysis dataset and condensed into time series of predictive indices. Provisional predictors were sought in four monthly grid point data fields: 200-mb u and υ winds, 1000-mb geopotential height (correlates > +0.98 with sea level pressure), and 500-mb geopotential height. All predictor types (composite, correlation, and residual fields) were used. It should be noted that all predictors selected were reanalysis “A” variables, which according to Kalnay et al. (1996) are primarily observation driven and therefore are mostly reliable. These predictors can also be reliably estimated in real-time applications for a forecast. The spatial domains of selected predictors were required to be of sufficient size so as to represent larger-scale phenomenon. The smallest area used was 10° × 15°. This size constraint reduced the possibility of selecting areas with purely random correlations as predictors.

A procedure used to further minimize areas of chance favorable covariation being chosen for predictors was to limit the predictor selection to the years 1958–1999. Selected predictors were tested against independent data during the first 9 yr (1949–57) of the database. The early years were thus utilized as an independent dataset to help determine if potential predictors were artifacts of chance correlation in the reanalysis data fields or were real representations of actual lagged physical teleconnection processes. If the possible predictor was similarly effective from 1949 to 1957 as 1958 to 1999, it was kept; otherwise, it was discarded.

All predictors that passed the initial test were candidates in the development of forecast equations from 51 yr of data. The predictors were selected using an all-subset technique allowing any predictor to combine with another predictor and these predictors were included in the forecast equations until the inclusion of any predictor explained less than 3% of additional total variance, as recommended by Gray et al. (1994), or until five predictors were obtained. No more than five predictors were selected to reduce the chance of overfitting. The correlation technique was an ordinary least sum of square deviations (OLS) regression scheme, unless this method produced negative forecast values. In these cases (H, IH) a Poisson regression was used as the correlation technique. The weights of the coefficients and predictors were selected to fit each of the nine dependent variables (NS, NSD, H, HD, IH, IHD, NTC, TONS, TOH) in maximizing hindcast skill.

After the equations were developed, the final step was performing a cross-validation (jackknife) procedure on the forecast models. The test, which was conducted following the guidelines suggested by Elsner and Schmertmann (1994), considered the extent to which each year could be predicted using data that were independent of the predictor observations for that year. Thus, cross validation in this paper consisted of predicting each year's TC activity for each of the 51 yr in the sample using parameters derived from the other 50 yr. This procedure partly emulates actual forecast conditions and provides some measure of the real predictive skill, generally regarded as an approximate upper boundary.

a. Galapagos and southeast Pacific predictors

The two most utilized predictors are the July 200-mb u and υ winds over the equatorial Pacific just west of South America. One or the other of these two predictors is involved in all of the TC forecast equations (see Table 5). When these areas experience winds that are anomalously westerly and southerly, August TC activity is generally suppressed in the Atlantic basin. The latter condition appears to be linked primarily to warm ENSO conditions, which produce anomalous westerlies in the deep tropical Atlantic (see Gray 1984b for a detailed explanation).

However, El Niño is not the only climate factor that affects the two “Galapagos” predictor values. There appears to be an influence from the Southern Hemisphere (SH) winter long-wave pattern. It is observed that SH troughs periodically extend to near the equator, creating upper-level westerly winds in the tropical eastern Pacific. These intrusions from the SH midlatitudes appear to be independent of the state of ENSO and tend to be short lived (on the order of 5–10 days) but can still radically change the prevailing 200-mb wind patterns and significantly alter u and υ flow anomalies (Fig. 6a).

During July prior to an active August, it is generally observed that SH upper-level westerly winds are at higher latitudes, replaced by a strong anticyclone over northern South America. This persistent upper high is characterized by easterly wind anomalies near the equator extending from the eastern coast of north Brazil westward to 120°W with northerly anomalies off the west coast of Ecuador, as shown in Fig. 6b. Such a pattern is favorable for Atlantic tropical-only TC genesis in a concurrent and predictive sense as it is associated with weak zonal wind shear and easterly 200-mb wind anomalies across the tropical Atlantic. It is hypothesized that conditions promoting the appearance and maintenance of this state of the atmosphere during July are a precursor signal to an active August. If this pattern recurs in August when the background climatology has become more favorable for TC formation, it often leads to a very active month.

Another predictor that is somewhat related to the equatorial winds is the SH midlatitude July 200-mb u wind west of Chile (40°–35°S, 110°–85°W) (region 4 in Fig. 5). When these upper-level westerly winds are weaker than normal, the duration of Atlantic TCs is typically increased, and consequently the NTC is increased. Although the equatorial 200-mb wind is only moderately correlated with the midlatitude flow (r = 0.25), using this midlatitude flow in combination with the equatorial wind significantly increases the amount of variance explained in HDs, IHDs, and NTCs.

b. Atlantic SLPA predictors

Two Atlantic Ocean SLPA areas are used as predictors. The most significant of these involves the July SLPA in the central Atlantic (25°–37.5°N, 47.5°–25°W; region 3 in Fig. 5). In general, low early season pressure in the Atlantic correlates well with increased Atlantic basin activity, a fact that has been known since the mid-1930s (Brennan 1935; Ray 1935). Anomalous low pressure is indicative of weak midtropospheric subsidence and less drying of the midatmosphere. Low pressure in the Atlantic Ocean is also associated with reduced trade winds, which is linked to warmer SSTs, partly due to decreased evaporation and upwelling.

April SLPAs in the equatorial Atlantic (10°S–5°N, 35°W–15°E) also display a strong inverse relationship with August TC activity, especially activity in the deep Tropics on both monthly (i.e., August) and seasonal time scales. Generally, if the April pressure in the equatorial Atlantic (region 10 in Fig. 5) is lower than normal, then lower August SLPA and negative zonal wind anomaly (ZWA) values are typically observed in the tropical Atlantic, correlating at r = 0.5. April SLPA may also be linked to ENSO wherein July SSTAs in Niño-4 correlate at r = 0.3 with this April index.

c. Greenland predictor

Another predictor from the Atlantic region is the June 200-mb zonal wind over the far northern region of Greenland (80°–85°N, 45°W–10°E; region 8 in Fig. 5). When the June 200-mb wind in this area is anomalously strong from the west, August TC activity is enhanced. This connection possibly reflects midlatitude blocking conditions near Greenland and, hence, the negative phase of the North Atlantic Oscillation (NAO). Enhanced ridging attending a negative NAO leads to easterly upper-level wind anomalies over northern midlatitudes while westerly anomalies occur over the polar latitudes. Van Loon and Rogers (1978) observed that enhanced wintertime blocking was more prevalent over the North Atlantic during the decades of the 1950s and 1960s than during the 1970s and 1980s. As TC activity during the 1950s and 1960s was much greater than during the 1970s and 1980s, it is hypothesized (Gray 2002a) in our 1 December seasonal forecast that the enhanced wintertime blocking patterns (implicit as westerly anomalies for the Greenland 200-mb zonal wind) are thus a harbinger of active hurricane seasons. Winter circulation patterns have become more similar to the 1950s and 1960s and correlate well with the recent spurt of increased seasonal and August Atlantic TC activity.

d. Northwest Pacific predictors

The Pacific region also holds some key predictors for August TC genesis. One of these concerns June SLPA in the northwest Pacific Ocean (18°–30°N, 134°–154°E), southeast of Japan (region 9 in Fig. 5). When June pressure anomalies in this region are high, August TC activity tends to be increased in the Atlantic basin. La Niña events are typically associated with high pressure in this predictor region (Larkin and Harrison 2001). This predictor could be linked to an observed tendency for significant reduction of TC activity in the northwest Pacific basin during June prior to active TC seasons in the Atlantic. In addition, this predictor is most closely linked to August IHD and HD explaining about 20%– 25% of the variance of these two parameters.

Sea level pressure anomalies over the Bering Sea region (47°–62°N, 156°E–164°W; region 2 in Fig. 5) during July are also strongly correlated with August Atlantic TC activity. There is more Atlantic TC activity when July pressure in this region is low. One of these two Pacific Ocean SLPA indicators (i.e., either June region 9 or July region 2) is selected by every forecast equation (except for TONS), emphasizing the linkages between Atlantic TC activity and the global circulation.

These Pacific region predictors suggest that important large-scale differences occur over the Pacific Ocean during summers prior to active versus inactive Augusts in the Atlantic basin. An idealized look at global features associated with an active pattern is shown in Fig. 7. In general, increased August Atlantic basin activity follows anomalous low pressure in the midlatitudes of the western Pacific Ocean and high pressure in the western tropical Pacific during the early summer. An effect of this pressure pattern is that vertical shear is typically greater in the tropical northwest Pacific Ocean during summers before active Atlantic years due to increased 200-mb westerly winds and 850-mb easterly winds, similar to what occurs during a La Niña event.

It is notable that a reverse (from the Pacific) synoptic-scale pattern is typically observed in the Atlantic Ocean in the early summer before an active August (Fig. 7). In particular, anomalous low pressure is noted in the Atlantic Tropics with a diminished Bermuda high and slightly increased pressures in higher latitudes. These lower pressures in the Atlantic subtropical high region are usually accompanied by low vertical wind shear in the main TC development region along with reduced trade winds and more easterly upper-level flow.

e. Southwest Pacific predictors

Another change related to August Atlantic TC activity occurs in the SH winter during July. In this instance, an index of 500-mb heights in the Southern Indian Ocean (SIO) (42.5°–57.5°S, 72.5°–95°E; region 5 in Fig. 5) is used as a predictor for IH and IHD. A weaker than normal ridge in the SIO is favorable for August TC activity in the Atlantic basin. The physical mechanism(s) for this relationship are not clear at present but may be linked to processes altering August ZWA-linked shear conditions across the eastern Atlantic. In this case, lower heights in the SIO are well correlated with easterly 200-mb zonal wind anomalies across the main Atlantic TC development region.

A predictor for August TC activity found in the SH is the July 200-mb u wind in the general vicinity of the Coral Sea (17.5°–7.5°S, 145°E–180°; region 6 in Fig. 5). When winds in this domain are enhanced from the west, Atlantic TC formation tends to be increased during the subsequent August. This predictor is also well correlated with ENSO, especially Niño-3.4 in the central Pacific as well as with increased zonal shear in the western Pacific. Anomalous westerly Coral Sea winds are closely tied to high pressure in the tropical central and eastern Pacific, as well as to cool water conditions in Niño-3.4 during August (hence La Niña).

f. Early season predictors

Additional TC predictors were found based on NH winter climate conditions. These include January sea level pressure over the southwest United States (30°– 40°N, 110°–95°W; region 12 in Fig. 5) and February sea level pressure over Scandinavia (52.5°–75°N, 5°W– 35°E; region 11 in Fig. 5), both of which are negatively correlated with subsequent TC activity. Low February sea level pressure in Scandinavia appears to be related to enhanced August TC activity through its association with stronger midlatitude blocking over Greenland. Enhanced blocking in the North Atlantic, and hence higher pressure near Greenland, are linked to a low pressure trough in the vicinity of Scandinavia.

The relationship between lower January pressure over the southwest United States and enhanced August TC activity is difficult to explain but could occur as a combination of an ongoing negative Pacific decadal oscillation (PDO) pattern (correlates at approximately 0.4 with southwest U.S. January pressure) and effects attending cold ENSO. Despite the lack of an obvious physical linkage, this January southwest SLPA association is very robust (see section 5) and needs to be explored further.

a. August landfall relationships

Other authors have attempted landfalling hurricane forecasts, most notably Lehmiller et al. (1997) and Elsner et al. (2000). These studies have shown some success in forecasting areas that could have a higher chance of a hurricane landfall than average. However, these forecasts were for the entire hurricane season and did not deal with month-to-month changes.

August TC landfalls in the United States are a relatively common occurrence. Table 10 gives a brief summary of statistics on U.S. August landfalls. Thirty-nine TCs have come ashore during the past 51 yr, an average of about one TC every 1.3 yr with about one H landfall during August every 2 yr and one IH every 5 yr. Multiple August hurricane strikes are a relatively rare event, with only six instances during the past 100 yr. In any given August, there is a 41% chance of an H landfall and a 10% chance of an IH landfall event. These probabilities could be adjusted based on a strong positive relationship between August NTC and the number of August TC landfalls on the U.S. coastline. Table 11 details this relationship for the years 1949–99 divided into two 15-yr subsections with vast differences in landfalling cyclones. Note that these relationships were very similar for observed and hindcast NTCs. Figure 9 compares the H and IH landfalls in the observed versus hindcast values for 15-yr periods. These strong contrasts between high and low NTC values allow for landfall probability forecasts based on forecast NTC. Monthly landfall probabilities allow for a more specific prediction than a broad-brushed seasonal forecast.

b. How August activity relates to total season activity

Relationships between August activity parameters and the seasonal totals are also considered. Table 12 displays a correlation matrix for August-only TC activity parameters versus total seasonal TC activity. The most reliable August indicators for total seasonal activity are HD and TOH. These parameters are influenced by generally long-lived hurricanes forming from tropical waves south of 20°N. When two or more TOHs occurred during August, seasonal NTC averaged more than 50% above normal, as shown in Table 13. Thus, the incidence of TOH during August is an important indicator for seasonal activity yet to come.

c. Test forecasts for August 2000–03

One of the few methods for diagnosing true forecast skill is to analyze the real-time performance of a technique. It is acknowledged that 3 yr of August TC forecasts are insufficient to determine forecast skill conclusively, but it may provide some useful insights into the future skill of the model.

Monthly forecasts for August of 2000–03 were issued at the beginning of August of these three years and discussed in the August seasonal forecast update produced by Gray et al. (2000, 2001, 2002b, 2003). The August forecasts included statistical and analog forecasts and a final adjusted forecast based upon input from both sources. The past 3 yr can be considered completely independent from the 1949–99 developmental database and are thus a useful test of forecast skill. It is important to note that the 2000 and 2001 forecasts were issued during the development of these statistical forecast schemes. The statistical technique utilized for the issuance of the 2000 and 2001 forecast (labeled experimental statistical forecast) has been changed. The final forecast scheme, which was developed with the same 1949–99 database (no input from 2000 or 2001), but which was not available at the time, is shown in Tables 14 and 15 as “final statistical forecast.”

Even during the development stage the forecasts verified reasonably well. Tables 14–Tables 17 show the 2000–03 statistical forecasts along with the qualitatively adjusted forecast and observed data. Forecasts for 2000 generally verified quite well and had skill well above climatology alone. The main disagreement between observed activity and the August 2001 forecast concerned whether an IH would occur during the month. The statistical models suggested there would be an intense hurricane while the analog analysis indicated only a small chance. This difference was the main cause of the error in the final August 2001 forecast. The statistical forecast for August 2002 was too high owing to favorable early season predictors. However, the final qualitative forecast took into account the rapidly changing climate conditions (early summer conditions very unfavorable due to ENSO) and resulted in a better forecast. The August 2003 forecast was excellent and was considered very successful.

The overall results for the test forecasts are very promising with indications of true forecasting skill. The statistical model is a useful guide for forecasting August monthly TC activity with additional subjective refinements further increasing skill. Further understanding of the underlying physical processes will lead to better forecasts, especially when extreme and unusual circumstances present themselves.