Literature review on the application of ARMA model in stock price prediction

University of Toronto, Faculty of Arts & Science, Toronto, ON, Canada

^* Corresponding author: messi.zhang@mail.utoronto.ca

Abstract

Predicting stock prices is a perennial quest in finance, yet price series are famously noisy, volatile and nonlinear. Among the many tools on offer, the autoregressive-moving-average (ARMA) model remains a surprisingly resilient workhorse. This review first sketches the logic of ARMA and the Box-Jenkins recipe for transforming raw prices into stationary returns, selecting lags and checking residuals. A broad body of evidence especially for 1 to 5 day horizons, confirms that ARMA delivers reliable, low-cost forecasts and, when paired with GARCH, can track volatility bursts with notable precision. Head-to-head studies show that on small samples or thin markets, ARMA often rivals much heavier deep-learning engines, while recent hybrids such as ARMA-LSTM and ARMA-Transformer marry linear transparency with nonlinear flexibility and shine on high-frequency data. We synthesise domestic and global findings, chart three clear trends, model fusion, finer time grids and AutoML optimisation, and flag the model’s blind spots: fixed coefficients, linear assumptions and sparse use of unstructured signals. Looking ahead, regime-switching ARMA, online updating, sentiment-rich inputs and risk-band forecasts (e.g., VaR, CVaR) promise to keep this classic framework both relevant and insightful.