1. Introduction
The method used by Struthwolf (1995) involved the use of 1200 UTC observed 850–700-mb thickness to predict the upcoming afternoon temperature at the U.S. Army Dugway Proving Ground in Utah. However, NWS forecasters always issue their first forecast for the upcoming afternoon before 1200 UTC and typically must use 0000 UTC model forecast data to produce the forecast. Nevertheless, since a linear relationship has been shown to exist between observed thickness and observed temperature, it is suggested that model thickness forecasts and observed surface temperature may exhibit a similar relationship, assuming the model forecasts to be reasonably accurate. Indeed, the collection of such data for a particular site during a period of time would likely aid in site-specific temperature forecasting, as well as provide useful information regarding model accuracy.
This study includes results based on data collected at the National Weather Service Office (NWSO) Nashville, Tennessee, during the 1-yr period of March 1995–96. Eta and Nested Grid Model (NGM) 24-h 1000–850-mb thickness forecasts for Nashville were collected and paired with the corresponding observed daily maximum temperatures. There were three initial objectives: 1) to determine if a correlation exists between low-level thickness forecasts and observed maximum temperatures, 2) to use mathematical regression methods to define any such correlations, and 3) to test the accuracy of the regression equations and assess their usefulness in forecasting maximum temperatures at Nashville.
A strong linear correlation was found between low- level thickness forecasts and observed maximum temperatures. Because of these findings, the goals of this study were expanded to include 4) four seasonal regression equations for the Eta Model and NGM and a comparison of model forecast accuracy (results showing the Eta Model to be more accurate), 5) a comparison of the eta regression equations to the NGM Model Output Statistics (MOS) forecasts to determine which numerical predictor was most accurate, 6) an additional study to determine if particular low-level relative humidity profiles and/or changes in such profiles were associated with large eta regression forecast errors (defined to be ≥3°F). It was discovered that certain changes in the average relative humidity of the 1000–850-mb layer at Nashville often occurred at times of poorest forecast accuracy.
Finally, an investigation was made to determine if eta regression forecasts could have been used during the study period to improve local temperature forecasts made at the Nashville NWSO. A discussion offered in section 5 reveals how the eta regression forecasts could, indeed, have been used in a subjective manner to adjust NGM MOS forecasts closer to the observed maximum temperature.
2. Methodology
a. Evaluation of data error sources and finding a“preferred model”
1) Estimation of point-specific thickness
Whenever forecasters use contoured graphics in the forecast process (such as those used in this study), the estimation of a point-specific thickness can represent a potential data error source. This error can be minimized by displaying a dot or gridpoint value on the map near the location where the thickness is to be estimated (i.e., Nashville, TN). This study used a 5-m interval in order to maintain simplicity and consistency, regardless of the season or thickness gradient (Fig. 1).
The Eta Model, produced by the National Centers for Environmental Prediction (NCEP), is available for distribution in two polar stereographic projections, 190.5- km resolution and 91.5-km resolution (true at 60°N latitude; Hoke 1987; E. Rogers 1997, personal communication) (Fig. 2). The Personal Computer Gridded Interactive Display and Diagnostic System (PCGRIDDS) eta graphics used in this study were obtained from the National Weather Service’s Southern Region Headquarters and utilized NCEP’s 190.5-km-resolution grid domain (B. Meisner 1997, personal communication). The PCGRIDDS NGM graphics used in this study were generated based on NCEP’s standard 91.5-km output grid domain.
2) Thickness forecast error and model reliability
Of the 142 days examined, only 7 low confidence cases were noted with the Eta Model, whereas 79 occurred with the NGM. Thus, it is suggested that confidence involving forecast accuracy calculations for the eta is greater than calculated using the NGM data. Conversely, due to Ei(v), it is suggested that when the NGM is involved, Eqs. (2) and (3) would often produce dissimilar results. As a result, it is unclear how much confidence can be ascribed to the calculations of NGM thickness forecast accuracy (when Tkr(v) is assumed to represent the true thickness). Indeed, compared to the eta, when using NGM data in Eq. (2), a higher percentage of absolute forecast error could typically result simply due to poor estimation of thickness.
With these results in mind, it is not surprising that calculations of forecast thickness error using Eq. (3) reveal that the Eta Model is typically most accurate (Table 1), not only on an annual basis but for three out of four seasons as well.
3) Model initialization error (at forecast time) and sounding error
As already mentioned, initialization error at verification time was computed for each day in order to determine the relative degree of confidence that might be ascribed to calculations of model thickness forecast error. However, it should be stated that no attempt was made to address model initialization errors that may have been introduced at the time the 24- or 12-h forecasts were made (0000 or 1200 UTC, forecast day), or the propagation of such error through the 24- and 12- h forecast cycles. Nevertheless, it is generally understood that a forecaster should be especially cautious when using forecast data from a model that fails to initialize well. Initialization error at forecast time should always be considered before introducing such forecast data into a forecast regression equation. If such error is observed, the forecaster can typically ascribe less confidence to the regression equation results. Finally, sounding instrumentation error was not addressed, and it is not known what impact such error might have contributed to the overall model thickness forecast error.
b. Evaluation of airmass changes
Only data from mostly sunny days were used in this study to minimize the effects of cloud cover and atmospheric moisture on daytime maximum temperatures. This included days when at least 65% of available sunshine was observed (using the standard NWS sunshine recorder). However, since certain low-level relative humidity profiles are often associated with certain temperature profiles, an investigation was made to determine if eta forecast errors might be associated with certain low-level relative humidity regimes. This involved calculating the average low-level relative humidity in the 1000–850-mb layer at verification time (0000 UTC, forecast day +1), as well as documenting changes in low-level relative humidity during the forecast period. To accomplish these objectives, the average 1000–850- mb relative humidity was calculated from each appropriate Nashville sounding (taken at forecast and verification times). Relative humidity is the only variable used to define an air mass in this study.
3. Seasonal regression forecast equations
Correlation ratios of three datasets must be quantified before describing regression forecast equations derived and utilized for this study. As expected, highest correlation (0.992) occurred between sounding observed thickness and maximum temperature. The eta (0.986) and NGM (0.985) correlations were virtually identical. These ratios, along with data plots in Figures 3a–c, indicate the relationship between forecast 1000–850-mb thickness and observed surface temperature is strongly linear and very similar to the relationship between observed thickness and observed surface temperature (Watson 1993). The high correlation ratios and strong linear relationships indicate that the use of ambient temperature as an approximation for mean layer virtual temperature is a reasonable assumption for this study.
Regression equations were derived for each season using 24-h thickness forecasts valid at 0000 UTC versus each day’s observed maximum temperature. The degree of forecast error generated by these equations is discussed below. Seasonal regression equations as well as the number of data days during each season are displayed in Table 2.
Comparisons of eta and NGM forecast thickness and observed thickness taken from atmospheric soundings with observed maximum temperatures show strong correlations (Figs. 3a–c). The regression lines in Figs. 3a and 3b were calculated using the last two equations in Table 2, respectively. Correlations readily support using regression equations derived from these data to aid in forecasting maximum temperatures on mostly sunny days.
As mentioned in section 2, the eta provided more accurate 1000–850-mb thickness forecasts during all seasons except summer, as well as on an overall annual basis. As a result, the eta also had the greatest temperature forecast accuracy (based on the derived regression equations) during the same time periods (when compared to the NGM; Table 1). Both models showed greatest forecast accuracy during the summer (when temperatures typically showed the least amount of variability). Correspondingly, less accurate forecasts occurred during the remaining seasons, when temperature variability was greater. This is not surprising when one considers the large thickness/temperature gradient that often exists over middle Tennessee during the coldest times of the year. The obvious sparsity of data apparent during winter was caused by a lack of mostly sunny days when percentage of available sunshine was at least 65%. Not surprising, among the three comparisons, the highest correlation is between observed thickness and observed maximum temperature.
It is interesting to note these regression equations indicate that a change of 3 m in the low-level thickness forecast correlates to a temperature change of ∼1°F. An increase by 3 m would imply about 1° of warming, and vice versa. However, since seasonal variability exists, four seasonal equations were derived rather than a single annual equation.
Twelve-hour forecast thicknesses were also taken from each 1200 UTC Eta Model to determine if it improved upon 24-h forecasts made 12 h earlier. Overall, 12-h forecasts improved by only about a meter (0.84 m), which is well within the PCGRIDDS estimation error. Separate regression equations were derived using 12-h forecast thicknesses. Their use improved maximum temperature forecasts only slightly.
4. Relationships between eta regression forecast errors and airmass types
Using the full dataset and the preferred eta regression equation, an attempt was made to determine whether a relationship existed between the eta regression forecast error and a “dry” or “moist” low-level atmosphere at forecast verification time. It was found that when average 1000–850-mb layer relative humidity was >50% (a moist day), the regression equation yielded a forecast temperature that was slightly too warm (average error = +0.5°F). Conversely, when average relative humidity was ≤50% (a dry day), the forecast temperature was slightly too cool (average error = −0.2°F). These small average forecast errors are well within the limit of error already allowed using thickness in the equation estimated from PCGRIDDS and are probably within range of typical temperature error of a sounding thermistor. Thus, no clear correlation between forecast accuracy and low-level moisture (at verification time) was observed. As a result, it was found that forecast accuracy would not have been significantly improved by introducing a“moisture bias variable” into the regression equations.
A similar investigation was made involving a more restricted 42-day subset of data in which forecast biases related to atmospheric moisture did appear. This data subset included only days when eta forecast errors were of sufficient magnitude to preclude the possibility of an error due only to PCGRIDDS estimation. As stated previously, a change in the 1000–850-mb thickness of 3 m typically corresponds to a temperature change of ∼1°F. Thus, a 3°F temperature change could be approximated by an equivalent thickness change of about 9 m. This is 4 m larger than the selected PCGRIDDS contour interval. As a result, by choosing a temperature forecast error of at least 3°F, one can reasonably expect the error to include additional sources other than those attributable to PCGRIDDS estimation. These sources might include propagation of an initialization error through the forecast cycle, an isolated (random) model error, or a model error of recurrent nature (i.e., a model bias). Using this subset, the average forecast error on moist and dry days was recalculated. The average error on moist days increased to +2.9°F while the average error on dry days was still less than 1° (−0.6°F).
The subset data were further utilized to investigate whether days with the largest forecast errors typically occurred in conjunction with some type of airmass change. To define airmass change, only relative humidity was used. This was done in order to isolate any model forecast biases that occur in conjunction with particular low-level moisture regimes and/or changes in the mean relative humidity of the 1000–850-mb layer. Average 1000–850-mb relative humidity was calculated from each appropriate Nashville sounding (taken at forecast and verification times). Finally, the change in average relative humidity during the forecast period (ΔRH = RH24 − RH0) was calculated. These data were available for 41 of the initial 42 days.
Results showed that during December–March, the Eta Model regression equation exhibited its greatest forecast error mostly on days when a significant change in low- level relative humidity occurred during the 24-h forecast period (where |ΔRH| ≥ 20%). Table 3 shows a summary of these results. A portion of this error might be attributed to difficulty in choosing a proper thickness value for use in the regression equation when a large thickness gradient existed over the forecast area. The environment associated with shallow cold fronts could also explain part of this error, where significant cold-air advection takes place only in the lowest levels, and the boundary layer is not well mixed. However, it could also indicate that the Eta Model cannot adequately forecast overall changes in the low-level environment (temperature and moisture content) under these conditions. Furthermore, the largest forecast errors almost always occurred on days when the air mass was transitioning to a drier regime (such as might be expected behind a cold front), often exhibiting a ΔRH of −20% to −70%.
The majority of days in the 41-day data subset exhibited a negative ΔRH, when considering RH24 − RH0. However, the associated ΔRH values were often not as large during the period April–November as those observed during December–March. Only 7 of 41 cases investigated indicated a positive ΔRH.
On days when a large temperature forecast error was accompanied by a significant change in low-level relative humidity (|ΔRH| ≥ 20%), the majority (86%) of temperature forecast errors were positive during May–February (Table 4). At such times, the eta regression equation would typically forecast maximum temperatures an average of 3.0°F too high. Furthermore, these positive errors almost always occurred when the air mass was transitioning to a drier regime. As previously mentioned, a shallow cold front moving through the area would be one situation that would produce an environment favorable for this type of error. Conversely, when large forecast errors and significant low-level relative humidity changes occurred during March and April, the eta regression forecast maximum temperatures an average of 1.3°F too low.
Positive temperature forecast errors that occurred on dry days (low-level RH < 50% at verification time) between May and February, which were also accompanied by significant drying (i.e., −20% ≤ ΔRH ≤ −70%), represent an anomaly in the data. (See Tables 4 and 5a.) Furthermore, a positive average forecast error during May–February was found to exist regardless of the size of error (Tables 5a and 5b). Negative forecast errors typically occurred on other dry days. For instance, days with significant drying and large forecast errors in March and April exhibited average forecast errors of −1.3°F (Table 5a). When considering all dry days with large forecast errors, average forecast errors equal −2.8°F. When this data is further divided into four subsets (based on the magnitude of ΔRH), the anomalous data (between May and February) becomes quite evident (Table 5a). In addition, if the anomalous data (days with large ΔRH between May and February) are eliminated from the 142-day record, average error for remaining dry days is −0.7°F.
Average forecast errors on moist days (i.e., low-level RH > 50% at verification time) with large forecast errors and significant low-level relative humidity change were +3.7°F. Even when these data are further divided into the five subsets listed in Table 5a, the average error for each subset remains positive. For moist days that exhibited a large forecast error and little or no moisture variance, average forecast errors were +2.4°F.
Positive temperature forecast errors that repeatedly occurred on moist days may offer additional insight into model bias. Since the Eta Model routinely exhibited a positive error bias on moist days (even when there was little or no change in moisture during the forecast period), perhaps the tendency to produce a positive error on certain dry days was also related to an eta tendency to forecast low-level environments that were too moist too long (even during times when a much drier air mass was about to enter the region). This possibility certainly warrants additional investigation and, as this study shows, days with significant low-level drying indicate forecast situations when eta regression (temperature) forecasts should probably be used with caution.
5. Comparisons of eta regression equation forecasts and NGM MOS
The question of when to use the maximum temperature forecast provided by the eta regression equation as opposed to NGM MOS guidance must be addressed. First, as shown in Fig. 4, NGM MOS typically fared quite well and provided a forecast maximum within 2°F of the observed maximum in 73% of 142 cases. In addition, comparisons of forecast accuracy between the eta regression and MOS are outlined in a summary of seasonal statistics provided in Table 6.
Based on the entire 142-day record, NGM MOS temperature forecasts were superior to the eta regression forecasts in about 50% of the cases when a significant change in low-level relative humidity occurred (ΔRH ≥ 20%). Furthermore, the greater the eta regression forecast error, the greater the likelihood that the MOS forecast was superior. For instance, at times when the eta regression forecast error was ≥3°F, MOS provided superior forecasts for 17 of 20 days. Of course, it is not surprising that a forecast scheme involving multiple predictors would often be more reliable than a forecast produced by a single regression equation at times when a significant change in low-level relative humidity occurs. However, it is not enough to simply look for the approach of a significant change in low-level relative humidity to determine when large eta regression forecast errors will occur or to determine the comparable accuracy of MOS forecasts. There were 15 cases when a significant change in low-level relative humidity occurred but was accompanied by eta regression forecast errors of less than 3°F. Furthermore, when a significant low-level relative humidity change was accompanied by forecast errors that were small (or zero), the eta regression forecasts were better than MOS almost half (46%) of the time. Unfortunately, the information collected during this study was insufficient to isolate the variable(s) that would help a forecaster know when a large ΔRH would be accompanied by a large eta regression forecast error or when the MOS forecast would be preferred.
Table 6 shows that MOS forecasts (during the 142 days of record) proved more accurate than the eta regression, most notably during the summer and slightly better during the autumn. During winter and spring, however, the eta regression proved more reliable. Therefore, when attempting to determine which forecast method to use, the particular season should be considered.
Forecasters are usually most interested (and perplexed) when forecast temperatures from different numerical models yield divergent results. Conversely, when numerical guidance is very similar, the forecast is usually less complicated, unless, of course, all models share the same error.
Using eta regression and MOS forecasts for all but the summer months, it is shown how eta forecasts might have been used to improve Nashville’s local forecast verification results. A review of these cases indicates the eta regression forecast would often have been useful in determining when positive or negative departures from MOS were most likely to occur.
There were 53 cases when MOS and eta regression forecasts differed by 3°F or more. However, since MOS was highly accurate during summer (Table 6), only the remaining three seasons were subjected to closer scrutiny to determine how eta forecasts might have been used to improve upon MOS. There were 39 cases during these three seasons when the difference between eta and MOS forecasts was 3°F or more. In 82% of these cases, eta regression indicated a proper “adjustment value” for utilization with MOS. Here, a proper adjustment value is defined as a numerical value that a forecaster can use subjectively to correct MOS closer to an observed maximum temperature. Hence, the eta regression would have directed the forecaster to make a positive correction to MOS on days when the MOS forecast was too low, and vice versa. In such cases it was found that the eta forecast would have typically been closer to the observed maximum temperature during spring and could have been used as the “best choice” forecast, whereas a MOS–eta average would have been the best choice forecast at most other times.
Fortunately, MOS data were often properly adjusted by the Nashville forecaster and local forecasts were issued that were in line with the eta adjustment factor 75% of the time. Nevertheless, if actual Eta regression forecasts had been used in the spring and an Eta –MOS average had been used at other times, 21 of Nashville’s 39 forecasts would have shown improvement, 12 of which would have improved by 2°–4°F. Eleven forecasts would have been worse, although 7 of these would have worsened by only 1°F.
Occurrences when Eta regression adjusted MOS properly by predicting maximum temperatures cooler than the NGM were clustered mostly during September through mid-October, and during the first half of March. Eta regression correctly predicted maximum temperatures higher than MOS most other times.
The eta regression forecast was least helpful (compared to MOS) when the difference between the two forecasts was only 1° or 2°F. During these times, the eta provided the wrong adjustment factor 57% of the time.
6. Conclusions
The purpose of this study has been to examine the correlation between 1000–850-mb forecast thickness and observed maximum temperature at Nashville, Tennessee. This was accomplished using statistical regression. It was shown that if low-level forecast thickness and temperature data are collected for a particular area or point, regression equations can be derived and utilized to aid in temperature forecasting.
Comparing the 1000–850-mb thickness forecasts between the Eta Model and the NGM, it was shown that the Eta Model low-level thickness forecasts for Nashville were generally more accurate than those produced by the NGM (based on comparisons with sounding data). Hence, the eta-derived temperature forecasts (using the regression equations) were usually more accurate as well.
It was also shown that dramatic changes in airmass moisture characteristics can have a significant effect on temperature forecasting. During December–March, most large eta regression forecast errors occurred during times when much drier air moved into the region.
A further analysis involving the eta regression forecasts and NGM MOS yielded more conclusive results regarding the use of regression forecasts as an aid in forecasting maximum temperatures. In Nashville, there are strong indications the seasonal eta regression equations discussed in this study can be used to help improve maximum temperature forecasts on mostly sunny days (using a MOS –Eta average), especially in autumn, spring, and winter, and at times when the eta regression forecast differs from MOS by at least 3°F. Also, there are times when the Eta Model’s superior low-level thickness forecast can be used as a singular predictor to forecast maximum surface temperature at Nashville with better accuracy than NGM MOS and its array of predictors (especially during the spring when the eta regression forecast differs from MOS by at least 3°F).
Compared to MOS, eta regression forecasts were least accurate during the summer and on about half the days when significant changes in average 1000–850-mb-layer relative humidity occurred (i.e., |ΔRHI| ≥ 20%). Overall, eta regression forecast errors exhibited a positive bias on moist days. A negative bias existed on dry days, except during periods of significant drying when a positive error bias occurred.
The method of temperature forecasting described in this paper is not meant to supersede any other method, but is rather intended to provide an additional aid that might be used, especially during mostly sunny days in which significant discrepancies exist between forecast guidance products.
Strengths in this method include the overall accuracy of eta forecast low-level thickness and the linear relationship between eta forecast low-level thickness and maximum temperature, from which a regression equation can be derived. Once an equation(s) has been derived, the only parameter required to produce a maximum temperature forecast is the forecast low-level thickness.
The method is reliable only during mostly sunny days, since effects of clouds and precipitation are usually not adequately assessed by low-level thickness. Thus, a forecaster must be confident that the upcoming day will be mostly sunny before attempting to use the regression equations shown in Table 2. Also, reliability of a maximum temperature forecast is dependent upon the accuracy of the low-level thickness forecast. Therefore, if model initialization is inaccurate by more than a few meters, results obtained using this method may be unreliable. Of course, model-initialized low-level thickness should always be checked for accuracy before using this method.
It must also be emphasized that model biases and forecast comparisons were calculated from a relatively small data sample (141 days). Additional data should be collected and preliminary conclusions should be tested using a larger number of days in order to ensure the results in this study are not transitory in nature.
It is hoped that future research regarding the use of atmospheric thickness will include developing techniques to include moisture parameters. This will help to eliminate biases caused by significant changes in airmass regime that are not accounted for in the low-level thickness forecasts.
Before relying upon this forecasting method at any station, local studies should be performed to establish a set of regression equations for that station. It is hoped that this paper has addressed (among many other items) a methodology for utilizing eta low-level thickness when forecasting maximum temperatures on mostly sunny days.
Acknowledgments
The authors thank Henry Steigerwaldt, SOO, NWSO Nashville, Tennessee, Dr. Steve Lyons, NHC; and Russell Pfost, SOO, NWSO Jackson, Mississippi; for their most helpful reviews and suggestions. The authors also thank Mark Richards, NWSO Nashville, Tennessee, for his assistance in collecting the data used in this study, Ron Flagstad, NWSO Nashville, Tennessee, for his helpful advice, and Eric Rogers, National Centers for Environmental Prediction, for his input regarding eta–NGM output grid domains. Special thanks are also given to the three anonymous reviewers for their insightful comments and suggestions for improving this manuscript.
REFERENCES
Hoke, J. E., 1987: Improving the horizontal resolution of the Nested Grid Model. NWS Technical Procedures Bull. 368.
Holton, J. R., 1992: An Introduction to Dynamic Meteorology. Academic Press.
Petersen, R. A., and J. M. Lord, 1996: Personal computer based gridded interactive display and diagnostic system (PCGRIDDS).
Struthwolf, M. E., 1995: Forecasting maximum temperatures through use of an adjusted 850- to 700-mb thickness technique. Wea. Forecasting,10, 160–171.
Watson, C. J., P. Billingsley, D. J. Croft, and D. V. Huntsberger, 1993:Statistics for Management and Economics. Allyn and Bacon.
Initialized PCGRIDDS 1000–850-mb thickness plot taken from 1200 UTC 4 Mar 1997 eta. The asterisk in middle Tennessee represents the location of Nashville.
Citation: Weather and Forecasting 12, 4; 10.1175/1520-0434(1997)012<0799:PDMTUL>2.0.CO;2
Eta Model output grid domains utilized by NCEP for PCGRIDDS. Note that the PCGRIDDS graphics utilized in this study were provided by the NWS Southern Region Headquarters and were based on the larger of the two domains (indicated by the bold outline), at 190.5-km resolution.
Citation: Weather and Forecasting 12, 4; 10.1175/1520-0434(1997)012<0799:PDMTUL>2.0.CO;2
(a) Eta forecast low-level thickness vs observed maximum temperature plot with the corresponding regression line (based on annual regression equation given in Table 1). (b) NGM forecast low- level thickness vs observed maximum temperature plot with the corresponding regression line (based on annual regression equation given in Table 1). (c) Observed sounding low-level thickness vs observed maximum temperature plot with the corresponding regression line. The data plotted in (a)–(c) represent all 142 cases used in this study.
Citation: Weather and Forecasting 12, 4; 10.1175/1520-0434(1997)012<0799:PDMTUL>2.0.CO;2
Performance of Eta –NGM MOS maximum temperature forecasts during the course of the study (where Eta forecasts are based on derived regression equations). The number of individual cases in which each forecast method differed from the observed maximum temperature for a certain departure (°F) is listed.
Citation: Weather and Forecasting 12, 4; 10.1175/1520-0434(1997)012<0799:PDMTUL>2.0.CO;2
Comparison of Eta –NGM thickness forecast and regression equation performance by season. Note that standard deviations are based on differences in each case between forecast thickness and observed thickness, as well as between maximum temperature forecasts and observed maximum temperatures.
Seasonal Eta –NGM regression equations derived from 0000 UTC 24-h model forecast thickness and the following afternoon’s observed maximum temperature. Note that “Tk” indicates model thickness (m), which is the input parameter. The equations yield forecast maximum temperature (°F).
Comparison of ΔRH for winter days when Eta regression forecast error was at least 3°F using a subset of those days shown in Table 1.
Comparison of ΔRH for days in May–February when ΔRH was at least 20% and forecast error was at least 3°F.
Table 5a. Eta regression forecast errors vs ΔRH for days with large errors (at least 3°F). Note the positive error anomaly for dry days during the May–February period. Note: There were 23 days on record for which temperature or relative humidity data were unavailable. Data for the remaining 119 days are listed in Tables 5a and 5b.
Table 5b. Eta regression forecast errors vs ΔRH for days with small errors (less than 3°F).
Comparison of eta regression performance vs MOS. The eta regression method proved more accurate overall; however, MOS proved more reliable during the summer and autumn. These statistics are based on data from 142 days.
