A Review On Breaking the Limits of Time Series Forecasting Algorithms

Abstract

Time Series (TS) forecasting has stagnated owing to algorithm restrictions, therefore systems developed using these methods can only perform so well. TS remains a challenge despite recent advances in Deep Learning (DL) in Natural Language Processing (NLP) and Reinforcement Learning (RL). This paper reviews the literature on these algorithms, highlights studies using them, and shows their limits. Neural Ordinary Differential Equations (NODEs) with continuous-time and continuous-depth tackle TS forecasting issues. Liquid Time-Constant (LTC) networks, a more advanced and reliable implementation of these NODEs, provides fluidity. We propose a new design that uses the LTC's liquid adaptability and is more adaptable to manage immediate changes. These algorithms are more steady, adaptive, and versatile than DL, which may help overcome its TS forecasting shortcomings.