FourierFlow Generator for Financial Time Series

The FourierFlow is a simple, yet powerful tool for generating synthetic time series data using deep generative models. The model is slightly modified and applied to the Market Scenario Generator Hackathon: From Stability to Storms (for more details of the hackathon, visit Hackathon Website).

Getting Started

Installation:

Clone this repository to your local machine. I recommend to create a new environment by running

conda env create -f environment.yml

Configuration:

• Navigate to the config_ff.yaml file.
• Adjust the parameters according to your needs:
  • input_dim: Dimensionality of input features.
  • output_dim: Dimensionality of output features.
  • hidden_dim: Size of hidden layers.
  • n_flows: Number of flow layers.
  • n_lags: Number of time lags.
  • vol_activation: Activation function for volatility modeling (e.g., “softplus”).
  • Other hyperparameters (batch size, learning rate, etc.).
• Pretraining or Checkpoint:
  • Choose between two modes:
    • Pretrain:
      • Load your training data and labels (regular and crisis data).
      • Train the generator using generator_regular.fit() and generator_crisis.fit().
      • Save the combined model dictionary using save_combined_model_dict().
    • Checkpoint (Pickle Files):
      • Load the pre-trained model from model_dict.pkl.
• Generating Samples:
  • Run the script main.py in console:

    python main.py

  • Specify the condition (e.g., crisis or regular) by setting condition[0] in the main.py.
  • The generated synthetic data will be saved to a pickle file.

References:

[1] Alaa, A.M., Chan, A.J., & Schaar, M.V. (2021). Generative Time-series Modeling with Fourier Flows. International Conference on Learning Representations.