A Study on Volatility Forecasting for the E Fund ChiNext ETF Using a CNN-LSTM Model

Abstract

Since its launch in 2009, the ChiNext Board has become a vital financing platform for high-growth small and medium-sized enterprises. The E Fund ChiNext ETF (159915), with its excellent liquidity, extremely low trading slippage, and complete data series, has become a core tool for investors participating in the ChiNext market. Accurate prediction of price volatility in this ETF carries considerable weight for short-term risk management and trading strategy refinement. Prior work has mostly relied on conventional time-series models or single deep learning structures to capture volatility, leaving the potential of hybrid deep learning approaches for such assets underexplored. To fill this gap, the present study develops a CNN-LSTM hybrid model: a convolutional neural network first extracts local temporal patterns, after which a long short-term memory network handles longer-range dependencies. Daily closing prices of the E Fund ChiNext ETF serve as the lone input, with the 5 day rolling annualized realized volatility designated as the target. Comparisons are drawn against the LSTM Attention model, the ARIMA(1,1,1) model, and the GARCH(1,1) model. On three core accuracy metrics—root mean square error, mean absolute error, and mean absolute percentage error—the CNN LSTM model outperforms all benchmarks by a clear margin. It also exhibits noticeably higher directional prediction accuracy, confirming that the hybrid architecture effectively grasps the intricate dynamics of financial time series. The central contribution of this work is the demonstration that a CNN LSTM model relying solely on price sequences is both applicable and superior for forecasting ChiNext ETF volatility. This offers market participants a short term risk monitoring tool that demands minimal data and is straightforward to implement, while simultaneously providing empirical guidance for model selection in analogous financial time series forecasting tasks.