| Stock forecasting is of great significance to many economists and investors,and to some extent,stock fluctuations reflect China’s economic level.Therefore,stock price forecasting has always been a focus of scholars’ attention.With the impact of the epidemic in recent years and changes in the international situation,the stock market has experienced significant fluctuations,and there are also increasing factors affecting the accuracy of prediction results.Neural networks can handle non-linear,multi-dimensional,and noisy data such as stocks well,and they also have good processing methods for long series data of stocks.However,neural networks also have the problem of high model computational complexity,so meta learning algorithms are introduced to address this issue,in order to use less data to achieve prediction results.Based on the above problems,this thesis establishes a MAML-LSTM Mixture model based on emotion analysis.The main process includes three aspects: data set establishment,model establishment and model comparison.(1)Firstly,establish a dataset.This article crawls the stock reviews of the corresponding stocks on the stock bar website,uses the Jieba algorithm for Chinese word segmentation,introduces stop words,removes meaningless vocabulary,and reduces subsequent workload.Using the K-means feature selection method,select positive and negative vocabulary,classify keywords,construct a financial domain sentiment dictionary,and calculate sentiment index;Then,use the Tushare interface to obtain historical stock data and calculate corresponding statistical indicators,combined with sentiment index,to construct a new dataset.(2)Secondly,establish a Long Short Memory Neural Network(Long Short Term Memory,LSTM)prediction model to predict stock volatility,and train the prediction model using Model Independent Meta Learning(MAML).MAML training LSTM to obtain an initial parameter with strong generalization ability θ,This initial parameter is θ The LSTM model can be fine-tuned in each subsequent task to quickly adapt to new tasks.(3)Apply the MAML-LSTM model to different training sets,and verify the accuracy of the model by dividing it into different datasets and using different stocks of the same type as the training set.The MAML-LSTM model was tested on three datasets: Ping An Bank,Yuyue Medical,and San An Optoelectronics.By comparing it with other methods,the impact of stock reviews on prediction results was verified.The data partitioning experiment combined with actual situations also reflected the impact of different partitioning methods on experimental results.The experimental results indicate that the dataset added to the stock review has a certain predictive effect on some sudden fluctuations,and the fitting effect of the predicted results using the new dataset partitioning method is better. |
