论文解读：Stock2Vec: A Hybrid Deep Learning Framework for Stock Market Prediction with Representation

1）Stock2Vec Embedding的框架：

• Feature:

Augmented features: MACD, PSAR, BB, SO, ROC, OBV, FI

Database "FinSentS Web News Sentiment": Contains the daily number of stock-related news articles along with sentiment scores for media texts. The dataset is accessible via https://www.quandl.com/databases/NS1/.

Seasonality features: month of year, day of month, day of week, etc.

Static features are static covariates such as the symbol name, sector, and industry category, for example.

• 将 feature 分为 categorical input 和 continuous input:

categorical feature：通过嵌入映射为 dense numerical vector, 股票名称中的向量被视为 category-specific Stock2Vec embeddings

continuous input：归一化到 0-1

• Feature benchmarking ：用 XBGBoost 选出得分前 20 的 feature

2**）The Hybrid Model：**

该预测模型采用了混合构建策略，并集成了Stock2Vec嵌入方法与Temporal Convolutional Network（TCN）。分析表明其输出并未直接生成单一大小为1的预测结果而是生成了一个向量作为特征图其中包含了从时间序列数据中提取的关键信息便于后续将其与已学习的Stock2Vec特征进行融合处理以提升模型性能

这里的TCN模块能够替代任何可学习时间序列模式架构的应用（包括LSTM、GRU等）。最后将一系列全连接层（称为"头层"）应用于特征组合以生成最终预测结果。

3**）Experiment：**

• Benchmark model：共7个

仅基于时间序列：TS-TCN and TS-LSTM

仅基于静态特征：random forest and XGBoost

Stock2Vec

混合模型：LSTM-Stock2Vec and TCN-Stock2Vec

• 评价指标：RMSE, MAE, MAPE, RMSPE