Anomaly Detection via Word Embeddings and Neural Networks

The price in the stock market is the result of the willingness of buyers and sellers to agree on a bid. Therefore, the most effective influences on stock prices are the investors directly involved in the stock transaction. On the contrary, the volatility of stock prices can also indirectly reflect the type of transactions at the current moment. Using this correspondence, the paper adopts the Skip-gram word embedding method in natural language processing to train the embeddings of various transaction types that may occur in the transaction process, and obtain the dense vector representation corresponding to the transaction types. After conducting experiments on multiple real stock datasets, the obtained experimental results show that the model trained with Skip-gram embedding has better prediction convergence effect. In addition, when the embedded results are used for abnormal stock price fluctuation detection task, they are more stable and obtain higher returns than the random embedding results or the non-embedding results.