Comparison of Time-Series Forecasting Models

Forecasting of time-series data has gained a lot of traction in modern times. Its application in all industries including healthcare, economics, supply chain, and so on makes it the go-to strategy for any business model or research application involving predictions. There have been many established and conventional statistical techniques implemented to effectively predict the next “lag” of time-series data including univariate Auto-regressive (AR), univariate Moving Average (MA), Simple Exponential Smoothing (SES), and more notably Auto-regressive Integrated Moving Average (ARIMA) along with its different variations [1]. Primarily, the ARIMA model has been observed to perform with the highest accuracy in the prediction of the next shift in the time-series. However, with the modern technological advancements which have led to a massive increase in the computational power of computers and the advent of machine learning approaches such as deep learning, the precision of the more orthodox models have been put under the scanner. Moreover, after realizing the power of time-series forecasting, many established organizations such as Facebook (with its FBProphet model) have also entered the fray. So naturally, an important research question arises; have modern deep-learning algorithms such as LSTMs taken over the more traditional approach namely the ARIMA model. Furthermore, do they even outperform the extensively researched FBProphet model specifically designed for forecasting? This research aims to unravel this mystery by using various performance metrics such as MAE, MAPE, and RMSE for each of these models. The models are tested to determine the Air Quality Index of Delhi, the capital of India.