Return-based factor models face two distinct lag problems. Both are inherent to inferring exposures from returns rather than observing them from holdings.
1. Style drift detection lag
The fundamental limitation: return-based models need enough return history to estimate exposures, so they can’t detect a manager’s style change until well after it happens.
Approaches
Fixed-window OLS (Sharpe 1992): the worst offender. A 36-month window means a sudden shift from value to momentum shows up gradually over 3 years as old data rolls off.
Exponentially-weighted regression (Two Sigma Venn): down-weights old observations with a decay function. A 3-year window with exponential weighting detects shifts faster than equal-weighted OLS because recent returns dominate. Still backward-looking.
Dynamic Style Analysis / Kalman filter (MPI Stylus): the most direct solution. Models betas as a state variable that evolves over time, using a Kalman filter to update exposures with each new observation. Detects style drift within weeks rather than months. The main methodological innovation of MPI’s DSA over standard RBSA.
Sequential shrinkage (Liquid Factor Models): takes a different angle. By using TIPLS estimation, betas are more stable to begin with, reducing false detection of “drift” that’s actually estimation noise. The trade-off: genuine drift is also detected more slowly.
The irreducible gap
None of these fully solve the problem. A manager who changes strategy overnight will still take time to show up in any return-based model. This is the core advantage of holdings-based models: they see the change immediately in the next holdings snapshot.
2. Cross-market lead-lag (asynchronous trading)
Different asset classes close at different times (Tokyo before London before New York). A global multi-asset model using daily close-to-close returns will underestimate correlations between assets that trade in different time zones.
Approaches
Newey-West adjustments (Two Sigma, Liquid Factor Models): correct standard errors for serial correlation induced by asynchronous trading.
Multi-day rolling returns (Two Sigma): uses 5-day rolling returns instead of single-day, capturing the full transmission of a shock across time zones.
L-day rolling windows / Dimson sum-of-betas (MAC3): addresses the related problem for market-beta estimation in multi-country models spanning multiple time zones.
Illiquidity-induced lag
For less liquid assets (private equity, small caps, high-yield bonds), reported returns lag true economic returns due to stale pricing. Rosenthal (2024) documents this directly: a high-yield bond fund has R-squared of 57% daily but 85% monthly, because the monthly horizon allows the lag to wash out. There is no clean fix other than using lower-frequency data or explicitly modeling the smoothing process.