为什么值得关注

该研究挑战了复杂模型泛化能力更强的普遍假设，提供了有力证据：将金融理论（简单锚点）嵌入模型，在域外场景下显著优于灵活的学习器，并表明域内准确性无法反映跨机制鲁棒性。

适合谁使用

金融机器学习研究者
量化分析师和波动率预测从业者
投资组合风险管理实践者
机器学习鲁棒性研究人员

典型使用场景

设计能在机制转换下保持技能鲁棒的波动率模型
通过在清晰分割的波动率机制上评估来压力测试预测模型
将机器学习模型的泛化能力与基于理论的基准进行比较
开发具有加性架构组件的物理信息神经网络

项目优势

清晰的实验设计，包含受控的机制转换和泄漏控制（42天隔离期）
可复现的代码、数据管道和共享CSV，确保所有模型使用完全相同的输入
六种模型的全面比较，并采用块自助法对R²差异进行置信区间估计
强有力的实证证据表明，跨机制泛化的关键在于结构而非容量

使用前须知

锚点的优势依赖于尾随的已观测波动率，因此它能在变化后快速调整，但无法预测变化
研究仅覆盖单一指数（KOSPI200），泛化到其他市场是未来工作
没有任何模型能预测机制变化，锚点只是减轻了跨机制失败，而非解决它

README 快速开始

Cross-Regime Robustness of Volatility Forecasting

Why a two-parameter mean-reversion anchor generalizes where XGBoost, Random Forests, and Transformers fail.

An empirical study of out-of-distribution robustness in volatility forecasting on the KOSPI200. Six models are compared on identical data under controlled regime transitions, testing whether learned models retain forecasting skill when the volatility level shifts — and whether embedding financial theory into a neural network repairs the failure.

Status: Undergraduate research project. Preprint + code archived for the record. If you use this work, please cite via the DOI in CITATION.cff (see below).

TL;DR — the main finding

Cross-regime volatility generalization is a problem of structure, not capacity. Flexible learners do not fail to find a signal; they discard transferable structure that a two-parameter formula retains, in exchange for regime-specific fit.

Cross-regime R² (train on one volatility regime, test on another):

Experiment	XGBoost	Random Forest	Transformer	soft-PINN	additive-PINN	anchor
E1 (low→high)	−1.15	−1.14	−1.11	−1.12	0.68	0.71
E2 (high→low)	−22.9	−27.2	−22.2	−10.1	0.70	0.76
E3 (mid→high)	−0.80	−0.78	−0.67	−0.76	0.59	0.57
E4 (mid→low)	−3.75	−5.77	−2.55	−4.41	−0.40	−0.96
E5 (bridge→mid)	−4.72	−10.5	−5.02	−3.47	0.32	0.69

All four data-driven models score negative cross-regime R² everywhere (Crisis Robustness Score = 0.00). A two-parameter closed-form mean-reversion anchor, with no training, achieves positive R² on four of five experiments.

Key results

Data-driven models: strong in-regime, zero cross-regime. XGBoost, Random Forest, and a Transformer all reach in-regime R² of 0.38–0.92 but negative R² on every regime transition.
In-regime accuracy is uninformative about cross-regime robustness. RF and XGBoost tie in-regime yet RF is strictly worse out-of-regime on every experiment — conventional validation cannot tell them apart.
Failure is directionally asymmetric. Downward extrapolation (crisis→calm) is catastrophic (R² −2.6 to −27.2); upward is bounded (−0.67 to −1.15). This replicates across three model

项目描述

Cross-Regime Robustness of Volatility Forecasting: Why a Two-Parameter Mean-Reversion Anchor Generalizes Where XGBoost, Random Forests, and Transformers Fail

volatility-regime-research