Application of Time Series Models in Stock Price Prediction
TL;DR
A hierarchical prediction framework of "baseline monitoring—volatility calibration—nonlinear correction," emphasizing that the true value of the model lies in probabilistic state reference rather than precise price prediction is proposed.
Abstract
Stock price prediction has long been a challenge in financial research due to the high noise, non stationarity, and weak predictability of the data. This paper uses daily market data of the CSI 300 Index from January 2020 to December 2025 as a sample. Under a rolling expansion window framework that strictly eliminates forward looking bias, it systematically compares the predictive performance of five types of time series models: random walk, ARIMA, GARCH, XGBoost, and LSTM. The results show that: random walk constitutes a formidable baseline in short term prediction; ARIMA has notable cost effectiveness in stable phases but underperforms when facing sudden fluctuations; the GARCH family of models has irreplaceable value in risk management; XGBoost has the best robustness in medium term prediction; and LSTM leads in multi step prediction but has limited economic relevance. Further testing by market phase reveals that simple models have the highest cost effectiveness in low volatility periods, while the advantages of nonlinear models only become apparent in high volatility periods. Based on this, this paper proposes a hierarchical prediction framework of "baseline monitoring—volatility calibration—nonlinear correction," emphasizing that the true value of the model lies in probabilistic state reference rather than precise price prediction.