NANSDE-Net: A Neural SDE Framework for Generating Time Series with Memory

Battle Count: 76

The paper directly addresses modeling financial time series with long- and short-memory characteristics (tested on SPX, TPX, SX5E equity indices). The ability to capture persistent dependence in returns and volatility is relevant for risk management, volatility forecasting, and algorithmic trading. However, the paper focuses on generative modeling rather than direct trading signal generation, and the model does not yet guarantee stationary increments which is important for financial modeling. The Itô-compatible framework makes it practical for derivative pricing and portfolio simulation.

Requires understanding of stochastic calculus (Itô processes, semimartingales, Volterra kernels), neural network parameterization of kernel functions, Markov augmentation techniques, and Euler-Maruyama numerical schemes. The training involves computing gradients through stochastic differential equations. However, the architecture itself is relatively simple (2-layer MLPs), and code is available. The theoretical foundations (existence/uniqueness proofs, backpropagation through SDEs) add significant complexity for full understanding.

Code is publicly available on GitHub (https://github.com/ozhr/ArmaNoise-SDE-Net). Experimental setup follows prior work [8] with clear hyperparameters (Adam optimizer, learning rate 0.004, 1000 iterations, 200 early stops). Network architecture (2-layer MLP with 20 hidden units) is specified. However, some details of the training procedure reference external papers for full specification.