N-beats Deep Learning Transformer Model For Nowcasting Consumer Price Index

Abstract

Accurate modelling of time-series data is vital across various domains, particularly in economic forecasting, such as predicting inflation rates. With inflation data typically released monthly, the limited number of observations poses a challenge for traditional modelling techniques. This study explores the applicability of the Neural Basis Expansion Analysis for Interpretable Time Series Forecasting (N-BEATS) transformer architecture to predict the Consumer Price Index (CPI). Transformers, commonly pre-trained on extensive datasets, offer promising capabilities for fine-tuning to specific tasks, even with limited data. In this research, we aim to replicate the N-BEATS transformer model architecture, utilizing monthly CPI data from the Kenya National Bureau of Statistics (KNBS). The analysis includes exploratory data analysis (EDA) to uncover patterns and trends, followed by model evaluation using Root Mean Squared Error (RMSE) and Mean Squared Error (MSE). This research endeavours to provide an alternative approach for inflation predictions to conventional deep learning and the traditional statistical modelling methods.