### Table 5.27: Tradeoffs in selection of a model set. Except for Logistic, all the model sets characterize well and predict as well as the data allows. Some of the model sets are more simple, but others have useful and interesting parameters.

1995

Cited by 10

### Table 1 are for a dynamic simulation for the 1954:1-1982:111 period. The results in Table 1 show, first of all, that the model fits the data well. The predicted values are based on a dynamic simulation of 115 quarters in length, and the root mean squared error over the entire period is only .Oll. Consider now the actual values in Table 1. There are two possible troughs that are rele-

"... In PAGE 5: ... Another is to see what the model predicts these values to be. This information is presented in Table1 . The model consists of ... In PAGE 6: ... For the last responses the trough might be in 1982, whereas for the earlier ones the trough is likely to be in 1980. Table1 shows that in 1980 the percentage of excess hours reached a high of 4.5 percent in the fourth quarter.... In PAGE 6: ... (In the aggregate model, other things being equal, the deeper the trough, the larger will be the predicted percentage of excess hours, and the companison of the two sets of results has not adjusted for different size troughs.) Second, the manufacturing sector may on average face deeper troughs than do other sectors, and the aggregate estimates in Table1 are for the total private sector, not just manufac- turing. One would thus expect the Medotf- Fay estimate to be somewhat higher than the aggregate estimates, and 8 percent versus a number around 5 percent seems consistent with this.... In PAGE 6: ... One would thus expect the Medotf- Fay estimate to be somewhat higher than the aggregate estimates, and 8 percent versus a number around 5 percent seems consistent with this. With respect to the predicted values in Table1 , in 1980 the predicted percentage of excess hours reached a high of 4.4 percent in the second quarter, and in 1982 it reached a high of 4.... ..."

### Table 2: Correlation coe cients The correlation coe cient is evaluated for each tem- plate with di erent value of T= . The template which gives the maximum correlation coe cient is declared as the matched model. Table 2 shows the correlation coe cients for the rough surface plotted in Figure 8. The correlation coe cient is found to be a maximum of 84% where the ratio T= is 2.5. Therefore the tem- plate with T= is declared as the matched model. It can be seen from Figure 11 that the prediction from the sonar model ts the experimental results very well. As predicted, the maximum amplitude drops and the acoustic energy spread wider as the sensor orientation increases.

### Table 1: Means and (Standard Deviations) of scores for several precedence functions Overall, the best results are obtained with the R2 statistic. In some sense, this is to be expected: we would like to give precedence to the links that predict well predicted variables. Although it seems as though more information about precedence could be obtained from pairs of variables, it is possible to construct the correct model by ordering the variables rather than the links. In addition, notice the low variance in the Wrong Not Reversed category. This e ect has a logical explanation: since most of the wrong links are discarded by the lter conditions, the di erences between these scores indicate links that were removed to avoid cycles.

1994

Cited by 4

### Table 3 shows that the simulation does extremely well at detecting when no error will occur (matching SWIFI for over 99% of the faults) and reasonably well at predicting the less severe injection results (e.g., the simulation correctly identi ed a dropped message for about 95% of the faults where SWIFI determined that result). The simulation, however, has relatively low accuracy in predicting severe fault results, such as host interface hang where about 20% of the injections match. One reason for the low level of accuracy is that the simulation does not fully model the interaction between the host and the interface. Several factors a ecting the accuracy level are addressed later in the discussion section.

1998

"... In PAGE 13: ...2% Total 385 385 83.1% Table3 : Breakdown of Number of Errors by Category for the Simulation The leftmost column of Table 3 shows the fault injection results (based on the categories in Table 2). The next column gives the number of faults resulting in each category observed in the simulations.... In PAGE 13: ...2% Total 385 385 83.1% Table 3: Breakdown of Number of Errors by Category for the Simulation The leftmost column of Table3 shows the fault injection results (based on the categories in Table 2). The next column gives the number of faults resulting in each category observed in the simulations.... ..."

Cited by 3

### Table 2. Automatically determined stressed words for dia- logue acts ACCEPT and SUGGEST. 3.1. Language Model Predictor The language model we use computes estimations for the occurrence of a word wi under the assumption of its pre- decessor words wi?1; : : : ; wi?n+1. To smooth the proba- bilities we combine the n-gram probability with smaller n-grams. There exist a lot of possible interpolation tech- niques for those probabilities { a very common one is the linear interpolation. Another interpolation method which performs quite well for the prediction task is the rational interpolation (cf. [7] for details) where the probabilistic model looks like this:

### Table 5.1 Univariate Regression Results In this table the results of the analysis described in the former chapters are depicted. In column three and four the expected dependence between the accounting ratios and the default probability is compared with the empirical relationship to default. As can be seen all variables behave in an economically meaningful way. In column five the functional form of the relationship between accounting ratio and log odd is listed, where l stands for linear, c for concave and u for an u-shaped empirical dependence. Besides, the univariate Accuracy Ratios for all data sets are shown. Notice that a powerful ratio for the default definition bankruptcy in general also does well in predicting default for the default criteria rescheduling and delay-in-payment. The variables that were selected for the backward selection procedure but that did not enter into the final model are marked with a star, while the winning accounting ratios are labeled with an exclamation mark.

### Table 5.1 continued Univariate Regression Results In this table the results of the analysis described in the former chapters are depicted. In column three and four the expected dependence between the accounting ratios and the default probability is compared with the empirical relationship to default. As can be seen all variables behave in an economically meaningful way. In column five the functional form of the relationship between accounting ratio and log odd is listed, where l stands for linear, c for concave and u for an u-shaped empirical dependence. Besides, the univariate Accuracy Ratios for all data sets are shown. Notice that a powerful ratio for the default definition bankruptcy in general also does well in predicting default for the default criteria rescheduling and delay-in-payment. The variables that were selected for the backward selection procedure but that did not enter into the final model are marked with a star, while the winning accounting ratios are labeled with an exclamation mark.

### (Table 5 in supplementary material). While dS was well predicted, the dN was overestimated. The observed ratio dN/dS is about twice as low as the predicted one.

### Tables 9 and 10 that frequency correlates fairly well with the predictions of the geome- try.18 Systems with inclusive persons are rarer than systems without, systems with dual number are rarer than systems without, and systems with paucal number are very rare indeed.

2002

Cited by 8