Comparing LSTM and Transformer Architectures for Daily Close Price Prediction on Chinese A-Share Stocks

DOI

10.2991/978-94-6239-801-6_62How to use a DOI?

Keywords

Quantitative Trading; Deep Learning; Time-series Forecasting; Stock Prediction

Abstract

Recent studies have reported strong performance of Transformer-based architectures for financial time-series prediction mainly evaluated on conventional error metrics. However, improvements in predictive accuracy do not necessarily translate into superior trading performance. This paper conducts a cross-sectional study to evaluate the performance of different financial forecasting models in the high-noise Chinese A-share market. In the study, a random-walk model, an Extreme Gradient Boosting (XGBoost) regressor, a two-layer Long Short Term Memory (LSTM), and a small encoder-only Transformer are trained in parallel to predict the next-day residual close prices using daily OHLCV data from 284 CSI 300 constituents over the period of 2015-2024. Approximately 2 million forecasts out of sample are generated under a strict time-series split. Besides, the models are then evaluated on various error metrics, and a long-only backtest with transaction costs taken into account. Results show that while all the models show limited predictive accuracy, their trading outcomes are significantly divergent. The Transformer achieves the best overall predictive performance, although its advantage over the random-walk benchmark is not statistically significant based on the Diebold-Mariano test. While the two-layer LSTM acts as a long-biased filter, achieving the highest market exposure and buy-and-hold returns. The findings suggest that the differences in trading performance are mainly attributed to different exposure behaviours induced by distinct models, rather than merely predictive accuracy. Therefore, it is significant to further evaluate financial forecasting models with both statistical criteria and multiple economic factors

Copyright

© 2026 The Author(s)

Open Access

Open Access This chapter is licensed under the terms of the Creative Commons Attribution-NonCommercial 4.0 International License (http://creativecommons.org/licenses/by-nc/4.0/), which permits any noncommercial use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license and indicate if changes were made.

Download article (PDF)

TY  - CONF
AU  - Zhiwei Zhang
PY  - 2026
DA  - 2026/10/23
TI  - Comparing LSTM and Transformer Architectures for Daily Close Price Prediction on Chinese A-Share Stocks
BT  - Proceedings of the 2026 2nd International Conference on Financial Innovation and Marketing Management (FIMM 2026)
PB  - Atlantis Press
SP  - 579
EP  - 589
SN  - 2352-5428
UR  - https://doi.org/10.2991/978-94-6239-801-6_62
DO  - 10.2991/978-94-6239-801-6_62
ID  - Zhang2026
ER  -