Models & Theory (模型与理论)
7 分类索引 — Models & Theory。完整概念清单在 concepts/index, 本页只列本分类。
行为金融 / 信号映射
- Degree of Rejoicing and Regret (DRR
- Ferson-Siegel Optimal Weighting (FS 权重
- Nonlinear Signal Mapping (非线性信号映射
时序与概率基础
- ARIMA Forecast Model (ARIMA 预测模型
- Random Walk Price Pathwise (参数化随机游走
- Filtration-Indexed Probability Space
- Filtered Probability Trend Hypothesis
- EM Algorithm for Factor Models
Lead-Lag / Directed Network (有向图 + 引线滞后)
Lead-Lag Detection (有向图 + pairwise 引线-滞后)
Cross-Correlation Functional (ccf-auc, ccf-lag1 (lead-lag 主流 metric)
Signature Lead-Lag Metric (rough path)
Distance Correlation (Székely 2007 (非线性 + 鲁棒相关)
XICOR — Chatterjee ξ Coefficient (Chatterjee 2019, 秩型 + 0↔独立 + 1↔函数 + O(n\log n))
MIC — Maximal Information Coefficient (Reshe 2019 Science, equitable correlation, Chatterjee 2019 反例证明非 function detector)
Directed Network Clustering (有向图聚类全景)
Hermitian Spectral Clustering (Cucuringu 2020, 直接优化 flow imbalance)
Meta-Flow Graph (聚类后可视化)
RowSum Leadingness (cluster lead/lag 排序)
Lead-Lag Equity Forecasting (Bennett 2022 Sharpe 0.62 信号)
Hermitian Laplacian (Laenen-Sun 2020, k-th 单位根 phase factor + Cheeger 风格保证)
Flow Ratio (chain objective (链式有向流目标, 仅相邻 cluster pair)
Spectral Sparsification (Spielman-Srivastava 2011, 加速至 nearly-linear)
MAX DICUT (有向 cut NP-hard, Goemans-Williamson 0.878 近似)
ML 架构
- Fisher-z Correlation Forecasting
- Graph Attention Network (GAT
- Hyper-Connections (超连接
- Residual Connection Tradeoffs
- 可微 Top-k 算子 (Differentiable Top-k Operator (sigmoid-shift + 一维二分; 精确保基数 \sum p_i = k; 对角减秩一 Jacobian 闭式 VJP)
- Label Horizon Paradox (标签视界悖论 — alias, Song-Liu-Chen 2026; ML 训练端 label design 范畴
- 前瞻回撤标签 (Forward Drawdown Label — 华安金工: 相对入场价 10 日最大跌幅作连续标签, 分布跨环境稳定
风险模型 / 高频
- 结构化高频风险模型 (Structured High-Frequency Risk Model) — 风险模型的分析对象从个股换成高频因子组合; 5 个基本面高频风险因子 + 5 个 PCA 统计风险因子; Adjusted R^2 约 75%
- 高频复合因子 (High-Frequency Composite Factor) — 按最大化夏普给高频因子配权、压成一个因子; 多空夏普 8.08; 加进中证1000 增强年化超额 +2-3pct
- 高频因子的四类信息 (Four Information Types) — 系列基础篇: 以「指标对时间重排敏不敏感」把高频指标分成分布/时序/关联/另类; 等权平均存在最优天数 n^*=20
- 高频因子的多空组合与有效性检验 (Long-Short Construction and Validity Tests) — 全市场按因子值配权替代 Top/Bottom 10%; 年化多空 >10% / 夏普 >3 / 回撤 <10% / 日换手 <40%; 时序相关性 <0.3 才算不同因子
- 高频因子风险识别 (Top-Down and Bottom-Up) — \lambda=\lambda_p+\lambda_r 拆出只提供波动的风险因子; 自上而下重建 Barra 式因子 + 自下而上时序相关性聚类; 截面对冲后 \rho^H\approx\frac{1}{1-\beta}\rho-1
因子与图结构
凸优化 / DRO / 共生
- ADMM 算法
- Marchenko-Pastur 分布与协方差病态
- Distributionally Robust Optimization
- Wasserstein Ambiguity Set
- 黎曼因子模型 (Riemannian Factor Model (Huang-RChen-YChen 2026: 切空间动态因子 + holonomy bound; BW 实现协方差 2 因子 ≈ LFM 9 因子; 因子 1 ≈ VIX)
- 切空间失配与和乐 bound (Tangent Space Mismatch and Holonomy Bound (parallel transport 对齐 + sine-theta 距离; 短记忆+强因子 → 维数无关 n^{-1/2})
- Covariance Shrinkage Estimation
- Stein Shrinker and Loss Functions
- James-Stein Shrinkage Estimation (Stein 现象 + JS 闭式 \lambda^\star = a/b + OS07 shrinkage t 根基)
- Graphical Gaussian Models (GGM, partial correlation, SS05 + LW 2003 闭式入口)
贝叶斯收缩 / 经验贝叶斯
- 经验贝叶斯 Beta-Binomial 收缩 — 样本越少越听市场的, 样本越多越信自己的; 矩估计先验: 用方差分解把股票间差异拆成「真实异质性」与「二项抽样噪声」, 噪声占比高则强收缩; \kappa 裁剪在 5–200; 放量事件率真实异质性只占 47.34%, 缩量占 61.53% —— 同一套先验对缩量有害
因子不确定性 / SV
- Factor Uncertainty Index (FUI
- JLN 不确定性测度框架
- Stochastic Volatility vs Deterministic Volatility
- Predictive Regression Diagnostics
跨资产网络 / Ψ
- 跨资产溢出矩阵与 SDF 估计
- 识别性与 T 比 N² 比率
- 正则化幻觉 vs 经济结构
- 可复制性 vs 经济增量价值
- 外生经济网络与跨资产预测
- 低秩交互矩阵 (Low-Rank Interaction Matrix
Embedding / 几何相似度
Micro / 微观母题
- Self-Fulfilling Impact Loop
- Volatility-Normalised Tick Size
- HFT Liquidity Withdrawal
- Order Flow Imbalance (stable alias)
- Price Impact
- Limit-Order vs Market-Order Execution
- LOB Geometry (Sparse vs Dense
状态检测
- Hidden Markov Regime Detection
- Wasserstein 模板追踪与标签切换修复
- Dynamic Regime Number Selection
- Bayesian Information Criterion (卡页)
- KNN Regime Estimator (卡页)
Volatility Forecasting / Crash
- HHIt
- Level-Independent Diffusion Baseline
- GARCH Miscalibration in Low-Vol Decile
- Realized Variance (RV
- GARCH(1,1
- HAR-RV
- Bipower Variation Jump Ratio
- Crash Forecasting (Crash-02
- Brier Score and Diebold-Mariano Test
- Logistic vs Normal-Tail Probability Mapping
- Fan-Wang 2007 Multi-Scale Jump Analysis
- Microstructure Noise
- Dirichlet Distribution
- VIX Term Structure
- Volatility Forecasting
Cross-ref
完整页清单
本文件夹全部 97 个概念页。返回: Concepts 总索引 · Quant Wiki 总索引
- 延迟套利与速度军备竞赛 — — 军备竞赛是市场设计的均衡常数: FTSE 100 上每分钟一次、中位数 46 微秒、占 21% 成交额; ES-SPY 相关性 10ms 0.1016 / 1ms 0.0080; 对策 = 频繁批量竞价
- 高频因子的四类信息 — — 高频因子按「信息来自哪里」分四类: 分布信息 / 时序信息 / 关联信息 / 另类信息; 判断一个统计量是不是「时序信息」的判据是 g=f(\mathrm{Reorder}(\mathrm{data}))\neq f(\mathrm{data}) —— 打乱日内顺序后会变; 这是整个兴业金工高频系列的分类骨架
- 高频因子的多空组合与有效性检验 — — 高频因子的换手远高于低频因子, 所以评价指标要跟着改: IC 半衰期、多空两分组按因子值加权 (不是常见的五分组等权)、IC 与 ICIR 的判读标准; 附截面相关性与时序相关性的分工
- 高频因子风险识别 — — 自上而下与自下而上两条建风险模型的路线; 区分「风险」「风险因子」「因子」三个概念; 高频因子的协方差矩阵不能直接用样本估计 (比 T 还多)
- 结构化高频风险模型 — — 把风险模型的分析对象从个股换成高频因子组合; 自上而下 5 个基本面高频风险因子 (逐次回归取残差保证正交) + 自下而上 5 个 PCA 统计风险因子; Adjusted R^2 均值约 75%, 强于纯 Barra 模型; 风险因子的 IC 移动均线应呈周期性 (mom_risk IC 均值 -0.001 / ICIR -0.005), 而非 Alpha 因子的长期为正
- 高频复合因子 — — 用风险模型估协方差, 再按最大化夏普给高频因子配权压成一个因子; 全市场 IC 0.100 / ICIR 1.364; 十分位 Top 组年化 49.82% (Barra 基准 38.51%), 多空夏普 8.08 (基准 5.91); 五个股票池内分位数严格单调; 加进中证1000 增强后年化超额提升 2-3 个百分点
- 经验贝叶斯 Beta-Binomial 收缩 — — 把个股命中率向市场先验收缩, 收缩强度由数据决定而非研究者设定; 矩估计 \kappa = m(1-m)/v^{between} (裁剪 5–200); 后验 \tilde p = w\cdot\hat p + (1-w)m, w = n/(n+\kappa); 放量/缩量的真实异质性占比 47.34% vs 61.53% —— 解释了为何贝叶斯对放量有效、对缩量有害
- ADMM 算法
- ARIMA Forecast Model
- Asset-Based Style Factors (ABS 因子
- Bipower Variation Jump Ratio
- Brier Score and Diebold-Mariano Test
- Covariance Shrinkage Estimation
- Crash Forecasting (Crash-02
- Cross-Correlation Functional (ccf-auc, ccf-lag1
- Degree-of-Rejoicing-and-Regret-(DRR
- Directed Network Clustering
- Dirichlet Distribution
- Distance Correlation (Székely 2007
- Distributionally Robust Optimization
- Dynamic Regime Number Selection
- EM Algorithm for Factor Models
- Factor Uncertainty Index (FUI) (因子不确定性指数
- Factor-Loading Uncertainty (因子载荷不确定性
- Fan-Wang 2007 Multi-Scale Jump Analysis
- Ferson-Siegel Optimal Weighting
- Filtered Probability Trend Hypothesis
- Filtration-Indexed Probability Space
- Fisher-z Correlation Forecasting
- Flow Ratio (chain objective
- Fung-Hsieh 七因子模型
- GARCH Miscalibration in Low-Vol Decile (Normal-Tail Assumption Failure
- GARCH(1,1
- Graph Attention Network (GAT
- Graphical Gaussian Models
- HAR-RV (Heterogeneous Autoregressive Realized Volatility
- HFT Liquidity Withdrawal (HFT 流动性撤回
- HHIt (Herfindahl Index of Intraday Variance
- Hermitian Laplacian
- Hermitian Spectral Clustering
- Hidden Markov Regime Detection
- Hyper-Connections
- Idiosyncratic Volatility Puzzle (特质波动率之谜
- JLN 不确定性测度框架 (Uncertainty as Conditional Volatility of Forecast Errors
- James-Stein Shrinkage Estimation
- KNN Regime Estimator
- LOB Geometry (Sparse vs Dense
- Lead-Lag Detection
- Lead-Lag Equity Forecasting
- Learning-Based Stochastic Volatility (学习驱动的随机波动率
- Level-Independent Diffusion Baseline (3n(n+2)
- Limit-Order vs Market-Order Execution
- Logistic vs Normal-Tail Probability Mapping
- Lookback Straddle Portfolio (回望跨式期权组合
- MAX DICUT
- MIC — Maximal Information Coefficient
- Marchenko-Pastur 分布与协方差病态
- Meta-Flow Graph
- Microstructure Noise
- Nonlinear Signal Mapping
- Order Flow Imbalance
- Ordinal Concordance (次序一致性
- Predictive Regression Diagnostics (预测回归的统计陷阱
- Price Impact
- Random Walk Price Pathwise
- Realized Variance (RV
- Recos (Rearrangement-Cosine Similarity
- Residual Connection Tradeoffs
- 可微 Top-k 算子 (Differentiable Top-k Operator
- Reverse-Recursive Residual Break Detection (逆时递归残差断点检测
- RowSum Leadingness
- Self-Fulfilling Impact Loop (自强化冲击环
- Signature Lead-Lag Metric
- Spectral Sparsification
- Stein Shrinker and Loss Functions
- Stochastic Volatility vs Deterministic Volatility (随机波动率 vs 确定性波动率
- VIX Term Structure
- Volatility Forecasting
- Volatility-Normalised Tick Size (波动率归一化 Tick Size
- Wasserstein Ambiguity Set
- 黎曼因子模型 (Riemannian Factor Model
- Panel Tree (P-Tree 目标导向面板聚类树
- Test Assets 与因子模型检验 (Test Assets for Factor Model Evaluation
- 切空间失配与和乐 bound (Tangent Space Mismatch and Holonomy Bound
- Wasserstein 模板追踪与标签切换修复
- XICOR — Chatterjee ξ Coefficient
- 低秩交互矩阵 (Low-Rank Interaction Matrix
- 可复制性 vs 经济增量价值
- 图诱导协方差 (Graph-Induced Covariance
- 外生经济网络与跨资产预测
- 暴露相似图 (Exposure-Similarity Graph
- 正则化幻觉 vs 经济结构
- 前瞻回撤标签 (Forward Drawdown Label
- 识别性与 T 比 N² 比率
- 跨资产溢出矩阵与 SDF 估计
- 日内收益的周期性 — — 同一半小时隔日延续, 显著至少 40 个交易日; 首日多空 +3.01 bp (t=9.62); 非 13 倍数滞后的反转全是买卖价差跳动; 成因未被解释
- 做市商的三重成本与碎片化 — — Menkveld 解剖单个 HFT 损益: 价差 +1.55 / 持仓 −0.68 欧元每笔; 持仓 <5 秒赚、>5 秒亏; Chi-X 参与率 8.1% vs Euronext 64.4%; 年化毛夏普 9.35