Predict Stock Returns using GloVe Embeddings and Document Vectors

One of my first projects at Apteo, the aim was to utilise different embeddings to improve the accuracy of neural network predictions. Apteo’s AI powered stock investing tool Milton uses diverse financial articles like company statements, latest news and other historical stock documents to extract important features from them. My task was to incorporate and test different embeddings to improve the model:

1. GloVe

2. Word2Vec

3. Doc2Vec

4. Average of GloVe and Word2Vec

From the above approaches, Doc2Vec was the most successful. This is because, when using GloVe or Word2Vec, word level features are created which has difficulty in capturing the overall sentiment of an article. Using document vectors, on the other hand, also helps compare 2 articles as a whole rather than break them into their corresponding words.

Note: Due to an NDA with Apteo, I cannot link the code.