This paper first investigates whether adopting the stationary bootstrap or the moving block bootstrap, instead of the usual independent bootstrap, in the random forest method improves forecasting of stationary time series. It is shown that the block bootstrap procedures adopted in the random forest method do not make any statistically significant improvement in Korean or US inflation forecasting. Secondly, we consider inflation forecasting in Korea using 93 macroeconomic/financial variables and various machine learning methods. The samples are from September 2004 to March 2022. Comparing total 13 models, one model outperforms the rest models for most forecast horizons, which is a simplified method of the model proposed by Kim and Han (2022). The method consists of the following two steps: 1) Select important variables based on the Boruta algorithm, 2) Using only those selected variables, implement the random forest and produce a forecast. The tests by Giacomini and White (2006) and Hansen et al. (2009) show that the model provides significantly better forecasts for most forest horizons. In particular, the Boruta algorithm selected total economically active population, total employed persons, BSI, house price as important variables for Korean inflation forecasting.