| Summary: | This paper explores the use of machine learning (ML) models in forecasting Singapore Consumer Price Index (CPI). Linear ML models such as LASSO, Ridge, Elastic Net and Support Vector Regression, as well as tree-based ML models such as AdaBoost, Bagged Decision Trees, Gradient Boosted Trees and Random Forests are evaluated on their predictive accuracy with root mean squared error as the evaluation metric. These models are also assessed against traditional forecasting models such as ARIMA and VAR which act as benchmark models. In addition, the model forecasts are separated into two sample periods to distinguish how periods of economic calm and heightened volatility such as COVID-19 could impact model performance. Our results showed that all four linear ML models were able to outperform the benchmark models during periods of economic calm, with LASSO at the forefront in terms of predictive accuracy. During periods of heightened volatility, results showed that model performance for the tree-based models notably improved, highlighting their ability to cope well with the higher variance. Nevertheless, among the suite of models evaluated, LASSO had the highest predictive accuracy across both sample periods, indicating its suitability for use in forecasting Singapore CPI. This could be attributed to Singapore’s exchange-rate-centred monetary policy which aptly deals with imported inflation, and the resulting low variance in the dataset could explain the better performance of linear ML models.
|
|---|