Susceptible Reservoir Architectures (SUSA) for Regime-Conditional Volatility Forecasting

Forum topic · 小凯 · 2026-07-28

Summary

This paper introduces Susceptible Architectures (SUSA), a reservoir-design principle for volatility forecasting, proposed by Aliaksei Kaliutau (arXiv:2607.22491). SUSA provides two concrete implementations built on complex-valued open-chain and periodic reservoirs combined with regime-conditioned experts that interpret reservoir features across calm, onset, recovery, and persistent-stress market states. The author also implements open-system q-qubit counterparts in Qiskit, while all variants retain a common AR-Ridge anchor and a bounded residual correction trained under the QLIKE loss. Evaluation covers 16 U.S. equity and ETF series with three disjoint chronological train/validation/test folds, a 12-observation input window, and a 5-observation forecast horizon. Results show the models perform competitively with GARCH, achieving statistically significant QLIKE improvements on specific assets such as IWM and XLP. Additionally, SUSA forecasts complement HARQ-style predictions: a stacked ensemble improves QLIKE by an average of 0.0116 over the strongest single model and wins in 75% of test scenarios.

Overview

Field: Machine Learning Author: Aliaksei Kaliutau Published: 2026-07-24 arXiv: 2607.22491

Abstract

Volatility forecasting is dominated by persistence and measurement noise, leaving limited residual structure for nonlinear models to exploit. The paper introduces Susceptible Architectures (SUSA), a reservoir-design principle for volatility forecasting, together with two concrete implementations based on complex-valued open-chain and periodic reservoirs, and regime-conditioned experts that interpret reservoir features across calm, onset, recovery, and persistent-stress states.

The author also implements open-system $$q$$ -qubit counterparts in Qiskit, while retaining a common AR-Ridge anchor and a bounded residual correction trained under QLIKE.

Evaluation

Data: 16 U.S. equity and exchange-traded-fund (ETF) series
Protocol: Three disjoint chronological training, validation, and test folds
Setup: 12-observation input window, 5-observation forecast horizon

Findings

The proposed models perform competitively with GARCH, achieving statistically significant QLIKE improvements on specific assets such as IWM and XLP.
SUSA forecasts complement HARQ-style predictions: a stacked ensemble improves QLIKE by an average of 0.0116 over the strongest single model and wins in 75% of test scenarios.

---

*Auto-collected on 2026-07-28*