Deep Learning and Dynamic Mode Decomposition for Inflation and Interest rate Forecasting

Abstract

Time- series forecasting in finance domain is a challenging task as financial data tends to be very volatile. In this paper, we forecast interest and inflation rates by comparing the performance of two forecasting methods, namely Long short term memory (LSTM) and Dynamic Mode Decomposition (DMD). LSTM and other Gated Recurrent Networks (GRUs) have proved their might in dealing with stochastic time series many times. However, DMD is a data-driven, spatio-temporal method that breaks down a system into spatial modes having varying temporal behaviour. Such modes assist us in evaluating how the system evolves and forecasting the system's future state. We then tune the hyper-parameters of the LSTM and DMD models using Bayesian Optimization (BO) technique. Eventually we compare the forecasts from both methods and tabulate the results.