Autoregressive Stock Value Prediction by Constructing a Novel Embedding Vector Space

Given a sequence of known stock "candles", can we predict the values of the next $n$ candles?

Transformer Model Approach

• A Transformer model is able to predict time series data that has complex nuanced relationships

  • For example, ChatGPT is a Transformer Model

• By predicting the next values $n$ times, appending each output to the time series, you can predict the $n$ next outcomes

  • These vectors are learned (trained) parameters that are optimized such that their inner product correlates to the ratio of the probability that those words appear near each other.

This is HUGE. By only knowing how many times 2 words were next to each other, the model learned who the CEO of Exxon is, the zip codes of Honolulu, and more!

• Graphic is visible in the attached Jupyter notebook.

Our Datathon submission focuses on constructing an embedding vector space that is able to carry this degree of high-level information about that particular stock, fueling our Autoregressive Transformer model to make strong predictions.

Check the attached link to see the full project presentation!

Built With

• python