股票是金融投资领域关注的热点，股票高风险高利润，流动性强

的特征普遍受到投资者的青睐。为了最大化收益，投资人需要对股票

价格进行预测来制定投资策略。随着机器学习与深度学习在量化领域

的应用越来越成熟，神经网络进行金融数据分析的算法得到广泛的应

用，而长短时记忆人工神经网络模型（LSTM）基于其长时间记忆与处

理数据的能力，在股票预测中具有良好的效果。但是，面对波动性较

大的数据时，模型寻找参数过程中可能出现的局部极值问题，以及过

拟合的问题。

本文主要通过构建贝叶斯优化的 LSTM 模型 BO-LSTM，对 LSTM

神经网络中的超参数寻找问题进行了优化，缓解了潜在的局部极值与

过拟合问题，然后以沪深市场上具有代表性的 4 支股票进行了实验预

测。首先，本文根据已有的股票研究成果选取影响股票波动的指标集，

并对指标集进行了数据清洗。其次，针对指标集相关性强的特征，本

文通过主成分分析法（PCA）对指标集进行了降维处理，生成的主成分

因子代替原有指标集作为神经网络的输入。然后，本文建立 BO-LSTM

模型，阐述了贝叶斯优化算法在 LSTM 神经网络模型中的原理以及理

论框架。最后，本文选取沪深市场中具有代表性的 4 支股票进行了预

测，并对 LSTM 以及 BO-LSTM 模型在数据预测方面的结果进行了评

价与分析。

结果表明，经过贝叶斯优化的 LSTM 神经网络模型在预测结果上

显著优于仅依赖梯度线性迭代的 LSTM 模型，预测的精度得到较大的

提升。平均精度提高了约 23%，最高精度提高了近 40%。

Stocks are a hot topic in the field of financial investment. Stocks, characterized by

high risk and high profit , as well as strong liquidity, are widely favored by investors. In

order to maximize returns, investors need to predict stock prices to formulate investment

strategies. With the increasing maturity of machine learning and deep learning applica

tions in the quantitative field, algorithms of neural networks for financial data analysis

have been widely applied. Long Short-Term Memory Artificial Neural Network models

(LSTM), based on their ability for long-term memory and data processing, have shown

good performance in stock prediction. However, it may encounter problems of local ex

tremum and overfitting during the process of finding parameters for data with high stock

volatility.

This paper mainly addresses the problem of finding hyperparameters in LSTM neural

networks by constructing a Bayesian Optimized LSTM model (BO-LSTM), which solves

the potential problems of local extremum and overfitting. The model is then tested on 4

representative stocks in the Shanghai and Shenzhen markets. First, this paper selects

a set of indicators that affect stock volatility based on existing stock research results,

and cleans the data for these indicators. Secondly, for features with strong correlation in

the indicator set, this paper conducts dimensionality reduction on the indicator set using

Principal Component Analysis (PCA), and the generated principal component factors are

used as inputs to the neural network. Then, this paper establishes the BO-LSTM model,

explaining the principle and theoretical framework of Bayesian optimization algorithm

in LSTM neural network models. Finally, this paper predicts 4 representative stocks in

the Shanghai and Shenzhen markets, and evaluates and analyzes the results of LSTM and

BO-LSTM models in terms of data prediction.

The results show that the LSTM neural network model optimized by Bayesian op

timization significantly outperforms the LSTM model relying solely on gradient linear

iteration in terms of prediction results, with a substantial increase in accuracy. The av

erage accuracy has improved by about 23%, and the highest accuracy has increased by

nearly 40%.