Sample: SPX option chains, 2018-01-02 to 2025-08-29, cleaned set 4,363,137 option rows across 1,926 trading dates; three standardized 30-day books studied: ATM straddle, 25d risk reversal, 25d/10d short put spread .
Main empirical result: uncalibrated LightGBM base quantile undercovers (exceedance 0.209 / 0.165 / 0.147 for the three books), while sequential conformal recalibration reduces exceedance to ~0.110 / 0.101 / 0.107, i.e. close to the 0.10 target and with lower violation magnitudes for most books .
Method and operational pipeline (base model + sequential conformal)
Base forecast:
Target: next-day normalized loss Y_{t+1} = (loss) / N_t, where N_t = gross option premium of option legs; forecasting unit is loss per unit of initial option premium .
Predictors: compact daily state vector (option-surface shape, chain-quality, market-wide indicators, short-run change vars) + book-level descriptors + lagged losses; rolling training window = 252 observations, retrain every 5 prediction dates; LightGBM quantile regressor as main learner (GBR/XGBoost robustness checks) .
Compute weighted upper empirical (1−α)-quantile of residuals B̂_t = Q_{1−α}^ω({R_s}); reported conformal threshold q̂^{conf}_t = q̂^{ref}_t + B̂_t. Floor (main spec) set to 0; warm-up/fallback logic: weighted buffer used once ≥30 residuals; fallback to unweighted quantile or most recent buffer or zero if necessary .
Dynamic stability: rolling 50-day worst exceedance reduced (e.g., ATM: 0.34→0.22; RR: 0.30→0.24; Put spread: 0.36→0.24) — conformal improves local tail-risk stability, especially where base forecast is most vulnerable .
Crisis window (2020-02-20 to 2020-04-15): conformal reduces crisis exceedance for ATM straddle (0.156→0.125) and put spread (0.206→0.147); risk reversal stays at 0.125 .
Robustness and operational notes
Robust across learners (LightGBM / GBR / XGBoost), marking rules (strict vs robust), and floor choices; conformal exceedance remains ≈0.10 across 15 spec combinations .
Next-day marking hierarchy (exact match → contract match → interpolation → nearest-neighbor fallback) matters for feasibility: robust marking increases usable sample and crisis feasibility but on the intersection sample it does not uniformly improve same-sample calibration — marking primarily expands operational coverage .
Marking distortion bound: |Y_{t+1} − Y^*{t+1}| ≤ (1/N_t) ∑ |w{t,ℓ}| ε_{t+1,ℓ}, with mode-specific ε bounds for interpolation/fallback .
思维导图
深度阅读
以下为对“Tenghan Zhong, Adaptive VaR Control for Standardized Option Books under Marking Frictions”报告的详尽逐章剖析与图表解读,结构化覆盖元数据、逐节精读、图表与表格逐一解构、估值/方法学解析、风险评估、批判性视角及最终综合结论。所有从原文获得的结论或推断在句末附有页码溯源标识以便追踪。[注:下文页码标识均对应上传文本的 page_idx 页面编号]
一、元数据与概览(引言与报告概览)
报告标题为“Adaptive VaR Control for Standardized Option Books under Marking Frictions”,作者为 Tenghan Zhong(University of Southern California),作者邮箱在文首列出。
标准化书的构造:三类书被确定并逐日按确定性规则选取合约与权重:30日ATM straddle、30日25-delta risk reversal、30d 25d/10d short put spread;若仅用期权腿存在残余方向性敞口,允许加入现货对冲以稳定暴露。此处强调“标准化构造+归一化”以确保时间可比性与经济解释性。
解读:Robust 标价显著扩充可评估样本(例如总体 n 从 strict 的较小值扩至 robust 的更大值),fallback 使用率中等,说明 robust 标价在实务层面提高了可操作性,但并不一定在同一日期上改善校准——文本中也明确区分了“可操作性价值 vs 同样样本的校准改进”。
Table 8(Additional mechanism decomposition by book,page 38-39)
描述:对每本书逐步列示 Base quantile only → Base + floor → Base + conformal → Robust final 的指标演变,重复验证作者对“conformal 为主要校正力”的结论(例如 ATM: Base 0.2048 → Base+conformal 0.1054 → Robust final 0.1105)。