Updated DTW+K-Means approach with LSTM and ARIMA-type models for Core Inflation forecasting

Abstract

The paper is dedicated to evaluating performance in forecasting tasks of the novel routine that includes adapted DTW + K-Means for aggregating series with similar dynamics. The algorithm was developed throughout the series of papers. Novel parts are designed in a way to work with periodic series, like in the investigated monthly data case. It is used over hundreds of Consumer Price Index components to find similar dynamics and aggregate them by the similarity of their dynamics. Then aggregated series are given as input to the ARIMA, SARIMA, and LSTM models, to forecast the total Core Consumer Price Index. The choice is based on the necessity to capture possible non-linear relationships between series. The dataset is quite rich and contains hundreds of Consumer Price Index components, which is a level of prices for different goods. Data suffers from multiple issues, including seasonality, so controlling them either with satellite models such as X-12 or with the architecture of the forecasting model is sufficient. The research results are important for different groups of agents. Private businesses seek to plan their pricing while government structures want to employ their administrative measures in a proactive data-driven manner. The result shows that the SARIMA currently outperforms other models. An LSTM model combined with DTW + K-Means method shows worse results yet it was able to catch non-linearities, unlike more traditional models. Further investigation of LSTM + DTW/K-Means performance and fitting is necessary.