This is one of the project tasks from the @GRIP Internship Program of The Spark Foundation
Predict the Indian Stock Exchange Sensitive Index - SENSEX
from historical stock price and news headlines data from 2015/01/01-31/03/2022.
- Transform Daily Close prices into logarithmic returns to attain better statistical properties, then used along side with Yesterday Close prices as numerical feature inputs.
- Utilize the Huggingface pretrained RoBERTa
cardiffnlp/twitter-roberta-base-sentiment-latest
model for Sentiment Analysis on news headlines. - The
LSTM
was trained on numerical data only and used as a Baseline to contrast with theLightGBM
which was trained on both numerical and textual analyzed data.
- The
LightGBM Regressor
model as expected was able to fit and generalize better than theLSTM
model with significantly lowerRMSE
(Root Mean Square Error). - The
LightGBM Classifier
was able to correctly predict the next day stock market's general movement (Buliish or Bearish) by 56.10%.