Regime-Dependent Predictive Structure Between Equity Factors: Evidence from Granger Causality Conference: ; ; CCS: Mathematics of computing Time series analysis CCS: Computing methodologies Causal reasoning and diagnostics

Chorok Lee · 2026

Read Regime-Dependent Predictive Structure Between Equity Factors: Evidence from Granger Causality Conference: ; ; CCS: Mathematics of computing Time series analysis CCS: Computing methodologies Causal reasoning and diagnostics by Chorok Lee (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-10732.txt.

Introduction

Chorok Lee documents how stock market factors predict each other in different market regimes. The work uses 35 years of Fama-French data, from 1990 to 2024. Factors sort stocks by traits like value and size. Lee tests whether one factor's moves help forecast another's, a pattern known as Granger causality.

The Value factor forecasts the Size factor during crisis periods, with a lag of nine days. The signal is strong in crises, with a p-value below one in ten thousand. In calm times, the link is absent. The split warns traders that factor ties seen in stress may vanish in normal markets, and the reverse may also hold.

Catalog record

CREATORChorok Lee
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE22 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2601-10732

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Regime-Dependent Predictive Structure Between Equity Factors: Evidence from Granger Causality Conference: ; ; CCS: Mathematics of computing Time series analysis CCS: Computing methodologies Causal reasoning and diagnostics Chorok Lee Abstract . We document regime-dependent predictive structure between equity factors. Using 35 years of Fama-French data (1990–2024), we find that the Value factor (High Minus Low, HML) Granger-causes the Size factor (Small Minus Big, SMB) during crisis regimes ( p < 10 − 4 p<10^{-4} , 9-day lag), but not during normal conditions. This pattern validates across 5 of 6 historical stress events (2008, 2011, 2015, 2018, 2020). We identify regimes using a Student- t t Hidden Markov Model (HMM), which detects moderate crises (2011: 69%) that Gaussian models miss entirely (0%). While this relationship does not generate trading profits, the 9-day lead time may inform risk management decisions. We emphasize that Granger causality establishes temporal precedence, not structural causality—common drivers could explain the pattern. Our economic interpretation is a hypothesis, not a verified mechanism. 1. Introduction The August 2007 quantitative meltdown, in which systematic equity strategies lost approximately 30% in three days ( Khandani and Lo 2011 ) , revealed a critical blind spot in factor-based risk management. When multiple quantitative funds held similar factor exposures, forced liquidation by one fund created price pressure that cascaded to all others. Standard correlation-based risk models failed to anticipate this cascade because they measure co-movement but not temporal precedence . The Missing Piece: Which Factor Moves First? Existing research establishes three stylized facts: (1) Factor correlations increase during market stress ( Ang and Chen 2002 ) ; (2) Returns exhibit regime-switching behavior ( Hamilton 1989 ) ; (3) Factor crowding amplifies drawdowns during liquidation ( Stein 2009 ) . However, a critical question remains underexplored: Does the predictive structure between factors change across market regimes? Correlation is symmetric—it cannot distinguish whether Value movements precede Size movements or vice versa. Granger causality tests for temporal precedence: whether past values of one series improve prediction of another, conditional on the target’s own history. If such predictive relationships vary by regime, we could: • Monitor the leading factor to anticipate movements in the lagging factor • Adjust hedges based on the current regime’s predictive structure • Detect regime transitions by observing when predictive links emerge or disappear Important caveat. Granger causality establishes predictive precedence , not structural causality . A statistically significant Granger relationship could arise from: (1) direct causal influence, (2) a common driver affecting both series with different lags, or (3) omitted variables. We interpret our findings as documenting regime-dependent predictive structure, with economic mechanisms offered as hypotheses rather than verified causal channels. 1.1. Our Findings Using Granger causality analysis within regime-dependent subsamples identified by a Student- t Hidden Markov Model, we establish: Main Finding: Crisis-regime predictive structure. The Value factor (HML) Granger-causes the Size factor (SMB) with a 9-day lag during crisis regimes ( p=1.89\times 10^{-5} ). This pattern validates out-of-sample across 5 of 6 historical stress events (2008, 2011, 2015, 2018, 2020). Normal regimes show no predictive link. Neither direction is significant, suggesting factors evolve more independently outside crisis periods. Practical Implication. During detected crisis regimes, Value factor movements precede Size movements by approximately 9 trading days, potentially providing lead time for hedging decisions. However, this predictive relationship does not translate to profitable trading (Section 5.2 ). 1.2. Contributions (1) Empirical Finding: First documentation that HML Granger-causes SMB specifically during crisis regimes, validated across 5 of 6 historical stress events (Section 4.3 ) (2) Methodological: Student- t HMM for regime detection that captures moderate crises missed by Gaussian models—accurate regime identification is prerequisite for discovering regime-dependent predictive structure (Sections 3.2 , 4.2 ) (3) Honest Assessment: Transparent evaluation showing the finding supports risk monitoring rather than alpha generation, with clear acknowledgment of what Granger causality can and cannot establish (Sections 5.2 , 5.4 ) 2. Related Work Factor crowding and systemic risk. Anton and Polk ( Anton and Polk 2014 ) show that stocks with common mutual fund ownership exhibit correlated returns. Lou and Polk ( Lou and Polk 2022 ) measure arbitrage activity through return comovement. Stein ( Stein 2009 ) formalizes how crowded trades amplify drawdowns. Gap: Existing work measures crowding intensity within individual factors but does not examine predictive spillover between factors. Regime-switching models in finance. Hamilton ( Hamilton 1989 ) introduced Markov-switching models for business cycles. Guidolin and Timmermann ( Guidolin and Timmermann 2007 ) extend regime models to asset allocation. Bulla ( Bulla 2011 ) applies Student- t HMMs to financial returns. Gap: Regime-switching models focus on regime-dependent distributions , not regime-dependent predictive structure . Granger causality in finance. Billio et al. ( Billio et al. 2012 ) construct Granger causality networks among financial institutions to measure systemic risk connectedness. Gap: No prior work examines regime-dependent Granger causality at the factor level. 3. Methodology 3.1. Data We use daily returns for the Fama-French six factors from Kenneth French’s data library: Market excess return (MKT-RF), Size (SMB), Value (HML), Profitability (RMW), Investment (CMA), and Momentum (MOM). Sample: January 2, 1990 – December 31, 2024 (8,817 trading days). 3.2. Student- t Hidden Markov Model Let z_{t}\in\{1,\ldots,K\} denote the latent regime at time t . We model: Transition model: P(z_{t}=k\mid z_{t-1}=j)=A_{jk} Emission model (multivariate Student- t ): (1) p(\mathbf{x}_{t}\mid z_{t}=k)\propto\left(1+\frac{\delta_{k}(\mathbf{x}_{t})}{\nu_{k}}\right)^{-\frac{\nu_{k}+d}{2}} where \delta_{k}(\mathbf{x}_{t})=(\mathbf{x}_{t}-\boldsymbol{\mu}_{k})^{\top}\boldsymbol{\Sigma}_{k}^{-1}(\mathbf{x}_{t}-\boldsymbol{\mu}_{k}) . Why Student- t ? Financial returns exhibit excess kurtosis. Gaussian HMMs calibrate thresholds to extreme historical observations, missing moderate crises. Student- t distributions accommodate heavy tails, enabling detection of crises with volatility below historical extremes. Model selection. The number of regimes ( K=3 ) was selected by the Bayesian Information Criterion, which favored three regimes over two ( \Delta BIC =847 ) or four ( \Delta BIC =312 ). Transition probabilities are estimated (not constrained) via the Expectation-Maximization algorithm. 3.3. Sample Overlap Consideration A methodological concern is that regime labels are identified using the full sample, then Granger tests are conducted within those regimes. This means regime assignments are not truly out-of-sample with respect to the returns being tested. We address this concern in two ways. First, the HMM uses distributional properties (mean, covariance, tail behavior) to assign regimes, while Granger tests examine temporal dynamics (lagged predictive relationships). These are distinct statistical features, limiting circularity. Second, our event-based validation (Section 5.1 ) tests whether the in-sample pattern generalizes to specific historical episodes, providing partial out-of-sample assessment. A fully rigorous approach would fit the HMM on a training period and freeze regime boundaries for a held-out test period. We did not pursue this because crisis regimes are rare (14% of sample), and splitting would leave insufficient crisis observations for reliable Granger estimation in either period. We acknowledge this as a limitation. 3.4. Per-Regime Granger Causality For each regime k , we extract observations: \mathcal{T}_{k}=\{t:\hat{z}_{t}=k\} . For each ordered pair of factors (i,j) , we test: H_{0}:r_{j,t}\perp\{r_{i,t-\ell}\}_{\ell=1}^{L}\mid\{r_{j,t-\ell}\}_{\ell=1}^{L} Significance threshold: \alpha=0.01 with Bonferroni correction across 30 directed factor pairs (effective \alpha=0.00033 ). Regime boundary handling. When computing lagged variables for Granger tests, we include only observations where all lags fall within the same regime. Specifically, for an observation at time t in regime k with maximum lag L , we require \hat{z}_{t-\ell}=k for all \ell\in\{1,\ldots,L\} . This excludes approximately 8%

Read the complete text