This paper investigates short-term inflation forecasting in Korea using a bottom-up machine learning framework. Price chages of Consumer Price Index (CPI) subcomponents are predicted with the Boruta–Random Forest algorithm and aggregated via expenditure weights. Forecasts are compared across aggregation schemes. The results demonstrate that disaggregated forecasting of 30 CPI subgroups followed by weighted aggregation achieves the highest accuracy, surpassing both aggregate forecasting approaches and survey-based benchmarks. Beyond improved predictive performance, the framework delivers detailed subgroup-level projections valuable for monetary policy and private-sector decisions, and it offers a scalable methodology for forecasting other macroeconomic indicators.