LSTM Stock Market Prediction

COMPSCI5103 Deep Learning For MSc (M) - 2024-25

This coursework applies deep learning techniques to forecast daily percentage movements in stock prices for 442 companies using historical market data from 2010 to 2022. The model leverages LSTM networks to capture temporal dependencies, with hyperparameter optimization via Optuna and model interpretability through Captum.

Coursework Overview

Objective: Predict stock price percentage movement for 442 companies on April 1st, 2022, based on historical daily movements from 05/04/2010 to 31/03/2022.

Jupyter Notebook

• Data Engineering:
  • Parsed and preprocessed historical stock percentage data.
  • Normalized and windowed time series for supervised learning.
• Model Development:
  • Implemented LSTM models using PyTorch.
  • Tuned number of layers, hidden units, learning rates, etc., via Optuna.
  • Final predictions generated for April 1st, 2022.
• Evaluation and Submission:
  • Model performance validated using Mean Squared Error (MSE).
  • Best model yielded a Kaggle public leaderboard MSE: 2.52854.
• Model Interpretation:
  • Captum used to evaluate feature importance in time series sequences.
  • Identified temporal patterns and influential time steps in prediction logic.

Selected Outputs

Optuna Optimization History Plot

• Efficient hyperparameter search led to significant MSE improvement.
• Tuned parameters: LSTM layers, hidden size, dropout rate, learning rate.

See: study_statistics.txt for detailed values.

Captum Attribution Analysis

• Temporal saliency maps identified recent days as most influential for predicting next-day movements.
• Interpretation revealed consistent model attention patterns across sectors.

Feedback

Excellent notebook structure and graphics for EDA, optimization and interpretation!

• You successfully split the original training dataset into a training and validation dataset so you could get an unbiased estimator of MSE loss - excellent. [score = 1 / 1]
• You did very good Optuna hyperparameter optimization to find optimal hyperparameters - excellent. [score = 1.0 / 1]
• You used the validation dataset in the Optuna optimization which is excellent. [score = 1.0 / 1]
• You did captum interpretation analysis where you discovered something about your model - excellent ! [score = 2 / 2]
• You had generally good code comments and markup including discussion sections of Notebook markup. [score = 2 / 2]

Notebook mark: 7.0 / 7

Kaggle MSE Score: 2.52854 which gives kaggle mark of 7 / 7

Final Overall Mark: 14.0 / 14 = 100%