Air pollution is one of the most critical environmental challenges facing modern urban centers, with the Air Quality Index (AQI) serving as a vital metric for public health management and municipal governance. This research provides an exhaustive investigation into the two primary paradigms of time-series forecasting for environmental data: the statistical Autoregressive Integrated Moving Average (ARIMA) and the deep learning-based Long Short-Term Memory (LSTM) network. While ARIMA offers a robust, interpretable framework for linear patterns through the Box-Jenkins methodology, LSTM architectures excel at capturing the complex, non- linear dependencies characteristic of atmospheric pollutants over extended sequences. Using a high-resolution dataset from the Central Pollution Control Board (CPCB) and meteorological ERA5 reanalysis for Delhi-NCR (2019–2025), we evaluate these models across diverse seasonal cycles. Results indicate that while ARIMA is computationally efficient for short-term, stable trends, the LSTM model achieves a significantly higher R^2 score (0.96) during extreme pollution events, such as the "Severe Plus" episode of November 2024 and the sustained hazardous spikes of October 2025. This paper aims to establish a benchmark for selecting the appropriate predictive architecture to support early-warning systems like the Graded Response Action Plan (GRAP).

 Keywords: Air Quality Index (AQI), Time Series Forecasting, Autoregressive Integrated Moving Average (ARIMA), Long Short-Term Memory (LSTM), Deep Learning, Machine Learning, Environmental Prediction, PM2.5, Urban Air Pollution, Delhi-NCR, ERA5 Reanalysis, Central Pollution Control Board (CPCB), Graded Response Action Plan (GRAP), Non-linear Modeling, Spatiotemporal Analysis.