Generating Financial Time Series by Matching Random Convolutional Features

Battle Count: 93

Highly relevant to quantitative trading. The paper directly addresses the core challenge of generating realistic financial time series from limited historical data, which is fundamental for backtesting, risk management, hedging, and strategy development. The conditional generation framework (generating futures given recent observations) mirrors how market simulators are deployed in practice. The evaluation includes PnL tail metrics from trading strategies, autocorrelation structure preservation, and cross-correlation modeling. The method outperforms signature-based and diffusion baselines on real financial datasets including stocks, FX, indices, and crypto. Applications span deep hedging, portfolio optimization, and stress testing.

Moderate complexity. SOCK itself is relatively straightforward: preprocessing (augmentation + normalization + random projection), grouped random convolutions with dilations, and softmax-based pooling. The pseudocode is ~50 lines of PyTorch. However, the full training pipeline involves: conditional GRU generator, feature matching loss with resampling schedule, multiple evaluation metrics (discriminative classifiers, distributional comparisons), and careful hyperparameter tuning. The generator architecture is simple (single-layer GRU), but training requires 100k steps (~1 hour on A40 GPU). The feature resampling schedule and augmentation choices require some tuning. Overall, a competent ML engineer could implement this in 1-2 weeks.

The paper provides detailed pseudocode for SOCK (Appendix B.4), generator architecture code (Appendix C.3.2), and randomized signature implementation. Hyperparameters are specified for all experiments. Datasets are described with sources (Yahoo Finance, Binance). However, no explicit GitHub repository URL is provided for the full codebase. The Diffusion-TS baseline references the authors' public implementation. Evaluation metrics are fully defined in Appendix C.2.