This research proposal examines the performance output of machine learning (ML) and deep learning (DL) models in predicting stock prices, with a focus on model architecture and training outcomes. Nine models, including traditional ML algorithms and novel DL architectures, are evaluated on a dataset comprising Indian stock market data. Notably, Linear Regression, Random Forest, and BiLSTM exhibit superior performance on both training and testing datasets, with low Mean Squared Error (MSE) scores and high predictive
accuracy (R²). Linear Regression achieves a train MSE of 20.59 and a test MSE of 16.21. Decision Tree achieves a train MSE close to zero but significantly higher test MSE at 50.90.
The Support Vector Machine shows high MSEs at 283,842 for training and 274,734 for testing, indicating poor generalization. Random Forest demonstrates competitive performance with a train MSE of 5.68 and a test MSE of 30.26. XGBoost, although performing well on training data (MSE: 5.50), struggles on unseen data (MSE: 141.47), indicating overfitting. Among the DL models, BiLSTM exhibits promising results with a train MSE of 8.22 and a test MSE of 0.64. Our proposed architecture, StockNet, shows encouraging performance with a train MSE of 8.06 and a test MSE of 0.07. However, traditional LSTM performs slightly worse, with a train MSE of 7.54 and a test MSE of 0.00. Lastly, the ResNLS Architecture performs well
with a train MSE of 0.00 and a test MSE of 0.07. These findings underscore the importance of model selection and architecture design in stock price prediction tasks and provide insights into the strengths and limitations of various ML and DL approaches in financial forecasting.