This project features a serverless machine learning pipeline for predicting the stock market price action based on past events. The resulting application can be found on: https://huggingface.co/spaces/tommyL99/Stock_Market_Prediction
This project is for educational and entertainment purposes only. It is not advised to base any real world financial decisions partly, or completely on this model.
The real life data source will be taken from Yahoo Finance using the yfinance package: https://pypi.org/project/yfinance/
For this application, it has been opted for a Long-Short-Term-Memory (LSTM) recurrent neural network (RNN) as this model has shown to achieve great performance on this exact task according to Moghar & Hamiche (https://www.sciencedirect.com/science/article/pii/S1877050920304865).
Based on this project's trained model, the following visual results were obtained based on AAPL stock, where MSE was determined at $8.23:
In the interface a user will be able to select the ticker symbol for a given stock in the list. Based on this, the app will use the yfinance package to retrieve up-to-date historical data and perform inference on this input. Advice will be displayed to the user based on the intent, i.e. a short or longer term investment.
As mentioned, the project is based on serverless or modular design, i.e. each pipeline is implemented separately to enhance testing; maintenance; and robustness of the system. Specifically this project consists of four main components:
This pipeline takes one or more stock symbols as input for which data is then retrieved through yfinance. This raw data is then processed to contain desired features in the correct data format in order to be interpreted by the training pipeline. These features are then saved using a feature store (Google Drive in this implementation). So this, or any application can access and retrieve the features on demand.
Using the feature store, the implemented LSTM model will be able to quickly access the data in the correct format. From there, the model is trained as usual, after which it is deployed on Hugging Face for easy access when doing inference.
Using the trained model accessible on Hugging Face, the inference pipeline simply parses the desired stock's raw data using the feature pipeline once more to get an interpretable format. Finally, the trained model is queried using these features to return a prediction.
In order to put this model into practice, an easily accessible UI is provided as a web application developed using Gradio. This application allows the user to select any stock from the list and input their intent in terms of the investment duration. From here, the model will interpret the prediction and make a verdict on the short term movement (positive or negative).