Graph-Neural-Network-for-Financial-Asset-Return-Prediction

OVERVIEW
--------------------------------------------------
Objective:
• Predict the future (t+1) excess returns for all assets (372 nodes) based on historical dynamic graphs.

--------------------------------------------------
Problem Context and Motivation:
Financial markets are inherently complex, with asset prices influenced by multiple interdependent factors such as supply chains, industry sectors, return similarities, volatility spillovers, and overall market conditions.
• Nodes: Each node represents a financial asset (372 assets total), characterized by 93 features (e.g., financial ratios, cash flow metrics, performance indicators). All assets are present at each time step.
• Edges: An edge is drawn between two assets if there is a long-term interrelationship between them. Edge weights are defined as the Pearson correlation of their historical excess returns over the past 36 months (ranging from -1 to 1). These connections are dynamic and change over time (e.g., Lehman Brothers removed in October 2008, Tesla added in June 2010, Apple's strategic pivot in 2019).

--------------------------------------------------
Project Goal:
Develop a Graph Neural Network (GNN) model that learns the evolving structure of financial networks and accurately predicts asset returns for the next time step (t+1). This model integrates both spatial information (asset interrelationships) and temporal dynamics (changes over time) to output a vector of predicted returns for all 372 assets for the upcoming month, providing a robust forecasting tool for financial decision-making.

--------------------------------------------------
Method:
A graph neural network (GNN) is employed with the following components:
• Graph Convolutional Layers (GCNConv): Extract node-level features from each graph.
• GRU Layer: Models the temporal dynamics across a sequence of graphs (historical time steps).
• Fully Connected Layer: Maps the GRU output to the predicted return for each asset.
• Dropout: Applied to the GCN outputs to prevent overfitting.

--------------------------------------------------
Data:
Training Set:
  - features_train.pkl – Shape: (175 time steps, 372 assets, 94 features)
  - graph_train.pkl – Shape: (175 time steps, 372 nodes, 372 nodes)

Test Set:
  - features_test.pkl – Shape: (33 time steps, 372 assets, 93 features)
  - graph_test.pkl – Shape: (33 time steps, 372 nodes, 372 nodes)

Note: Each training sample includes asset features (with the first column as the label for excess return) and a corresponding graph structure. The test set does not include labels.

--------------------------------------------------
Model Architecture:
TGCN Model:
• Graph Convolutional Layers:
  - A list of GCNConv layers is applied over the historical time steps to extract spatial features for each asset.
• GRU Layer:
  - The output from the GCN layers (collected over the historical window) is fed into a GRU layer to capture temporal dependencies.
• Fully Connected Layer:
  - The final hidden state from the GRU is passed through a linear layer to predict the excess return for each asset.

--------------------------------------------------
Training Details:
Data Preprocessing:
• Asset features are preprocessed using scaling (StandardScaler, MinMaxScaler, or RobustScaler) and optionally reduced via PCA.
• Adjacency matrices are normalized if specified.

Dataset Splitting:
• The data is split into training and validation sets based on time steps.

Training Loop:
• Loss: Mean Squared Error (MSE) between predicted and actual returns.
• Optimization: Adam optimizer with weight decay.
• Learning Rate Scheduler: ReduceLROnPlateau is used to adjust the learning rate based on validation loss.
• Early Stopping: Training stops if validation loss does not improve over a specified number of epochs.

Testing:
• After training, the best model is loaded to generate predictions for the test set.
• Predicted returns for all assets are saved in a CSV file for submission.

--------------------------------------------------
Code Structure:
Notebook:
• DL_ap2938_Week4.ipynb - Colab version containing the full implementation, including:
  - Data loading and preprocessing.
  - Definition of the AssetDataset and AssetBatch classes.
  - Implementation of the TGCNModel class.
  - Training, validation, and testing loops.
  - Generation of the submission CSV file.

Dependencies:
• torch_geometric for graph neural network operations.
• easydict, numpy, pandas, pickle, scipy, and sklearn for data processing.
• torch and torch.nn for building and training the model.

--------------------------------------------------
Conclusion:
This project demonstrates the application of graph neural networks combined with recurrent layers to capture both spatial and temporal dynamics in financial networks. By processing historical graphs representing asset interdependencies, the model predicts future asset returns, which can be evaluated using RMSE. All code, experiments, and results are provided in this repository.