Diachronic Sample Integration: Robust Tail-Risk Estimation with Generative Models

Battle Count: 68

Highly relevant to quantitative trading. The paper directly addresses tail-risk estimation (VaR, ES) for trading strategies using generative simulators, which is central to risk management in algorithmic trading. The fixed-budget simulation framework mirrors practical constraints in trading (compute, latency, validation resources). The evaluation uses 65 canonical trading strategies (Buy-and-Hold, Mean Reversion, Trend Following) on high-frequency Nasdaq data. The method improves calibration of risk measures critical for position sizing, capital allocation, and regulatory compliance (Basel III/FRTB). The approach is architecture-agnostic and requires no retraining, making it immediately deployable in existing trading simulation pipelines.

Low implementation complexity. DSI is a test-time-only method requiring no modification to the generative model architecture, training objective, or additional training runs. The core algorithm (Algorithm 1) simply involves: (1) selecting spaced checkpoints from a training trajectory, (2) drawing samples from each checkpoint, (3) pooling samples, and (4) computing tail functionals on the pooled empirical distribution. The main practical considerations are choosing appropriate burn-in point, mixture size K, and stride M. No new model training or architecture changes are needed.

Source code is publicly available on GitHub. Synthetic data generation procedure is fully specified with exact parameters. Market data (Nasdaq ITCH) is a well-known public dataset. Training details (3000 epochs, AdamW, lr=1e-5, U-Net architecture) are provided. Results averaged over 10 independent runs. However, exact random seeds and full hyperparameter sweeps for all baselines may require additional detail.