Patch-VTT: A Memory-Efficient Transformer for Volatility-Based Anomaly Detection and Risk-Off Portfolio Management

Detecting market stress at daily resolution is central to modern risk modeling, yet the deep architectures capable of it are often too heavy to be practical. The Variable-Temporal Transformer (VTT) captures inter-variable structure effectively [1], but its attention scales with the raw sequence length, inflating model size and destabilizing threshold calibration on the small samples typical of financial data. We introduce Patch-VTT, which tokenizes each input window into patches [2] so that attention operates over compact regime-level tokens rather than individual days. Using S&P 500 daily data from 1980 to 2026, we build five volatility features from a GARCH(1,1) model [3] and reconstruct 100-day windows through a patched variable-temporal backbone, combined with a cross-attention head that restores per-day resolution and a rank-based fusion of the patch and point reconstruction errors. Days whose fused error exceeds the 95th percentile of the validation distribution are flagged. The patched design cuts the backbone parameter count by roughly thirtyfold and its dominant attention dimension by close to sevenfold, while yielding detections stable across random initializations. The model recovers the principal stress episodes of 2018 to 2025, and most flagged days correspond to innovational volatility-regime shifts rather than isolated shocks [4]. A risk-off overlay that reduces equity exposure following each flag raises the Sharpe ratio from 0.62 to 0.89±0.22 and lowers the maximum drawdown from 34% to 19% relative to buy-and-hold; the improvement persists once the 2020 crisis is excluded (Sharpe ratio 0.91±0.26). Patch tokenization thus makes transformer-based volatility detection lighter and steadier, with period-robust economic value.

References:

Kang, H., & Kang, P. (2024). Transformer-based multivariate time series anomaly detection using inter-variable attention mechanism. Knowledge-Based Systems, 290, 111507.
Nie, Y., Nguyen, N. H., Sinha, P., & Kalagnanam, J. (2023). A time series is worth 64 words: Long-term forecasting with transformers. ICLR.
Bollerslev, T. (1986). Generalized autoregressive conditional heteroskedasticity. Journal of Econometrics, 31(3), 307–327.
Chen, C., & Liu, L. M. (1993). Joint estimation of model parameters and outlier effects in time series. JASA, 88(421), 284–297.