On business cycle forecasting

Frontiers of Business Research in China

Table 4 Out-of-sample estimation results: AUC

Forecast horizon (months)^a	1	3	6	9	12
Models	AUC
Model with channels (a), (b), and (c) simultaneously^b
AR-Logit-Factor-MIDAS	0.913 (0.000)	0.937 (0.000)	0.953 (0.000)	0.925 (0.000)	0.927 (0.000)
Models with either channel (a) or (b)
ST-Probit-MCF	0.930 (0.000)	0.929 (0.000)	0.933 (0.000)	0.926 (0.000)	0.933 (0.000)
AR-Probit-YS	0.877 (0.000)	0.880 (0.000)	0.889 (0.000)	0.901 (0.000)	0.906 (0.000)
Models without any channel (a), (b) and (c)
ST-Probit-YS-EI	0.910 (0.000)	0.893 (0.000)	0.889 (0.000)	0.901 (0.000)	0.897 (0.000)
ST-Probit-YS	0.857 (0.000)	0.842 (0.000)	0.876 (0.000)	0.864 (0.000)	0.866 (0.000)

^a For each forecast horizon N (where N = 1, 3, 6, 9, 12), we evaluate the ability of each model to forecast the probability of a recession within the next N months
^bChannel (a): using a flexible function form by including a lagged recession probability into a dynamic Logit or Probit model of recession forecasting; Channel (b): employing a dynamic factor model to extract common factors from many monthly or weekly economic and financial variables; Channel (c): applying MIDAS to incorporate the mixed-frequency common factors in the dynamic Logit framework