Forecasting daily emergency department arrivals using high-dimensional multivariate data: a feature selection approach

Table 2 Model accuracies in terms of absolute percentage errors

	Mean	Standard deviation	Median	Max	Differs from SN (p)	Worse than best (p)
Naive	8.4	6.4	6.9	36.4	1.00	< 0.001
ARIMAX-A	8.4	6.2	6.9	33.7	1.00	< 0.001
RLS-U	8.3	6.2	7.1	37.7	1.00	< 0.001
SNaive	8.2	6.6	6.6	41.8		< 0.001
ARIMAX-SA	8.0	6.5	6.5	39.0	1.00	< 0.001
RF-FS	8.0	5.9	6.6	33.5	1.00	0.002
LMS-FS	7.8	5.9	6.5	32.6	0.98	0.007
RF-SA	7.7	5.7	6.5	28.5	0.72	0.035
RF-U	7.5	5.7	6.1	33.2	0.42	0.10
RF-A	7.4	5.7	6.4	36.6	0.22	0.22
LMS-A	7.3	5.6	6.3	34.3	0.16	0.30
ARIMAX-FS	7.3	5.9	5.9	36.2	0.12	0.37
LMS-SA	7.2	5.5	6.1	31.6	0.07	0.53
RLS-A	7.2	5.5	6.4	39.3	0.048	0.64
ARIMA	7.1	5.5	5.7	29.5	0.019	0.86
LMS-U	7.0	5.3	5.8	30.7	0.011	0.95
RLS-SA	6.9	5.1	5.9	24.6	0.003	1.00
RLS-FS	6.9	5.2	5.9	30.1	0.002	1.00
ARIMAX-W	6.6	5.3	5.3	31.7	< 0.001

ARIMA autoregressive integrated moving average, ARIMAX regression with ARIMA errors, RLS recursive least squares, RF random forest, LMS least mean squares, SA simulated annealing, FS floating search, SNaive = seasonal naïve, A all features, U univariate, W Whitt’s features. Statistical significance is calculated using two-tailed ANOVA with Dunnet’s post hoc test for multiple comparisons

ISSN: 1472-6947