Exploring an LSTM-SARIMA routine for core inflation forecasting

Анотація

The object of the research is the Core Inflation Forecasting. The paper investigates the performance of the novel model routine in the exercise of the Core Inflation Forecasting. It aggregates 300+ components into 6 by the similarity of their dynamics using an updated DTW algorithm fine-tuned for monthly time series and the K-Means algorithm for grouping. Then the SARIMA model extracts linear and seasonal components, which is followed by an LSTM model that captures non-linearities and interdependencies. It solves the problem of high-quality inflation forecasting using a disaggregated dataset. While standard and traditional econometric techniques are focused on the limited sets of data that consists just a couple of variables, proposed methodology is able to capture richer part of the volatility comprising more information. The model is compared with a huge pool of other models, simple ones like Random Walk and SARIMA, to ML models like XGBoost, Random Forest and simple LSTM. While all Data Science model shows decent performance, the DTW+K-Means+SARIMA+LSTM routine gives the best RMSE over 1-month ahead and 2-month ahead forecasts, which proves the high quality of the proposed forecasting model and solves the key problem of the paper. It is explained by the model's capability to capture both linear/seasonal patterns from the data using SARIMA part as long as it non-linear and interdependent using LSTM approach. Models are fitted for the case of Ukraine as long as they’ve been estimated on the corresponding data and may be actively used for further inflation forecasting.

Класифікація

Ідентифікатори

Схожі роботи