Skip to content

Stock price prediction using a Temporal Fusion Transformer

Notifications You must be signed in to change notification settings

Soham-Deshpande/Stock-TFT

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Temporal Fusion Transformer - NEA

A Technical Indicator for liquid asset valuation forecasts using a Temporal Fusion Transformer

The model I will be exploring is a transformer-based deep learning architecture that takes advantage of attention, more specifically multi-head attention in my implementation. Many models use ARIMA(Auto-Regressive Integrated Moving Average) but I am proposing using transformers with multi-head attention, something that I will talk about later on, to help the model ‘learn’ about the stock rather than just trying to fit a curve on it based on the last few data points. The benefits of learning allow the model to consider previous experience instead of just looking at a few, previous data points. This technique will hopefully result in a higher accuracy compared to ARIMA. Other implementations have managed an accuracy of 94%(Zolkepli and Divino, n.d.) so I will be aiming to stay within 5%. The reduced target comes down to a few factors such as limited access hardware, not enough time to optimise my program for the hardware as well a few other factors.

The project write up can be read here.

Use intructions: 1. pip install -r requirements.txt 2. run PytorchForecasting.py