Table 2.

Median prediction interval width at confidence levels from 10 to 99%

Epsilon	Nonconformity measure	Confidence level
		10%	20%	30%	40%	50%	60%	70%	80%	90%	95%	99%
${10}^{-3}$	Abs-diff	0.109	0.221	0.336	0.462	0.604	0.771	0.986	1.284	1.813	2.237	3.841
	Normalized	0.122	0.243	0.362	0.478	0.595	0.718	0.854	1.027	1.319	1.649	2.892
	Log-normalized, $\beta =0$	0.071	0.155	0.257	0.387	0.560	0.801	1.171	1.812	3.291	5.273	10.879
	Log-normalized, $\beta =1$	0.074	0.159	0.260	0.384	0.545	0.763	1.080	1.599	2.689	4.031	7.676
${10}^{-4}$	Abs-diff	0.069	0.139	0.211	0.288	0.378	0.486	0.629	0.843	1.245	1.695	3.006
	Normalized	0.079	0.157	0.233	0.311	0.395	0.491	0.610	0.789	1.200	1.918	7.194
	Log-normalized, $\beta =0$	0.042	0.094	0.159	0.243	0.352	0.519	0.772	1.223	2.311	3.918	10.157
	Log-normalized, $\beta =1$	0.044	0.097	0.163	0.245	0.356	0.509	0.741	1.137	2.030	3.233	7.204
${10}^{-5}$	Abs-diff	0.065	0.132	0.201	0.270	0.354	0.459	0.600	0.813	1.217	1.680	3.024
	Normalized	0.075	0.148	0.220	0.293	0.376	0.474	0.605	0.824	1.445	2.664	12.199
	Log-normalized, $\beta =0$	0.041	0.092	0.155	0.234	0.341	0.495	0.738	1.171	2.205	3.747	10.007
	Log-normalized, $\beta =1$	0.042	0.095	0.158	0.235	0.339	0.486	0.710	1.095	1.963	3.156	7.247

Shown are MPI at confidence levels (validity) from 10 to 99%. Note that a smaller median prediction interval indicates higher efficiency of a nonconformity measure. Shown are results for models with $\mathit{cost}=1$ and epsilon values ${10}^{-3}$, ${10}^{-4}$ and ${10}^{-5}$. Italicized are results for the best model at each epsilon value and confidence level. Marked by bolditalics are results for overall best models at each confidence level