Table 7 shows the results of a test consisting of 2l6 predictions 

 and hindcasts at stations BRAVO, CHARLIE, and DELTA. A prediction was 

 made for each day on which at least 3 BT's were available for verifica- 

 tion. Predictions were scattered randomly over several years for all 

 months of the seasonal thermocline. Corrections of stability index for 

 salinity were not possible because salinity gradients were not known. 

 BRAVO and DELTA are located in permanent currents, where salinity gradi- 

 ents and their variations may often be very large. Prediction conditions 

 at these locations can be considered very unfavorable, while prediction 

 conditions at CHARLIE are rather favorable. 



Although Table 7 gives a general view of the test, it is somewhat 

 distorted, because the prediction error in feet for shallow thermoclines 

 is a large percentage of the observed depth; whereas the same error in 

 feet for deep thermoclines is a small percentage of the observed depth. 



The data used for compiling Table 7 were also used to compute the 

 mean values of the observed thermocline depth and of the prediction error 

 for each station of the test (Table 8). 



As mentioned previously, ECHO (35° N, kd° W) seems to lie in an area 

 of permanent convergence during late summer and autumn. This area is not 

 convergent during spring and early summer. A test of 21 predictions was 

 made at this station in April, May, June, and July. The results of this 

 test are included in Table 8. Distribution of negative and positive errors 

 of prediction at BRAVO, CHARLIE, DELTA, and ECHO does not indicate varia- 

 bility of thermocline depth with latitude. 



Figure 36A shows the frequency distribution of absolute values of pre- 

 diction errors in percent of the mean observed thermocline depth. This 

 figure was constructed from the combined data of Table 7 and hindcasts for 

 ECHO during spring and early summer. Figure 36B shows the frequency dis- 

 tribution of the same data at CHARLIE and ECHO only. Nine hindcasts with 

 widely scattered errors exceeding 50 percent of the observed thermocline 

 depth were excluded from the frequency distribution in Figure 3&A, although 

 they are included in Table 7. No such prediction errors occurred at CHARLIE 

 and ECHO. The 9 omitted cases were made in areas of permanent currents and 

 were all positive. Some of these errors were apparently caused by failure 

 to recognize convergent or divergent effects during prediction. The error 

 was less than 10 percent of the observed thermocline depth 50 percent of 

 the time in Figure 3&A and was less than 10 percent of the observed value 

 65 percent of the time in Figure 36B. 



Figure 3&C shows the frequency distribution of negative and positive 

 errors from the observed values of the same data in Figure 3&A. The mean 

 of the errors ? is -1.8 percent. The negative shift is mainly caused by 

 excessive negative errors at CHARLIE. Since data of CHARE IE were used to 

 develop this prediction method, a closer balance between negative and posi- 

 tive errors than would occur at the other stations can be expected. Exami- 

 nation of weather data for hindcasts revealed a general tendency to under- 

 estimate the rate of convergence when hindcasting thermocline depth. Fre- 

 quent convergence in the area of CHARLIE is quite strong, because water 



73 



