Liquid Factor Models, proposed by Rosenthal (2024), construct factor models exclusively from liquid, cheaply tradeable instruments (futures, government bonds, liquid ETFs) rather than from long-short characteristic-sorted portfolios or holdings-based decompositions. The model achieves explanatory power comparable to Fama-French models while producing more stable betas and cutting hedging costs from hundreds of basis points to single digits.
The 10 factors
| Factor | Proxy | Category |
|---|---|---|
| SP500.xs | S&P 500 excess return over 3M T-bill | Equity |
| R2KSPX | Russell 2000 minus S&P 500 (size) | Equity |
| R1KVMG | Russell 1000 Value minus Growth (value) | Equity |
| YCLevel | Average yield change across 3M-30Y USTs | Yield curve |
| YCSlope | Weighted 30Y-3M and 10Y-2Y yield spreads (zero DV01) | Yield curve |
| YCBump | Second-derivative approximation of yield curve | Yield curve |
| IG10 | iBoxx IG return minus 10Y UST (zero duration) | Credit |
| HY5IG10 | (iBoxx HY minus 5Y UST) minus IG10 | Credit |
| InflSurp5Y | Change in 5Y TIPS breakeven | Other |
| VIX | Change in VIX | Other |
Key criteria for factor selection: transparent, actively traded, narrow bid-ask spreads, accessible via listed derivatives, minimal rebalancing.
Methodology: Theory-Informed Partial Least Squares (TIPLS)
Factors are grouped into blocks (equity, yield curve, credit, other) and estimated sequentially in order of relevance to the portfolio’s dominant asset class. Each block’s residuals become the next block’s dependent variable. This is equivalent to Type I ANOVA and acts as a shrinkage estimator, trading unbiasedness for lower variance.
Before estimation, raw variables are combined into correlation-reducing pairs (e.g., A and B-A instead of A and B) to reduce multicollinearity. This works when the original correlation exceeds 0.5. Variance inflation factors drop from 64 (kitchen sink) to under 5.3.
Key empirical results
- Berkshire Hathaway: R-squared 74% (vs. 76% for Fama-French-Carhart 6-factor), revealing value exposure plus yield curve sensitivity that FF models miss entirely. No significant alpha.
- High-yield bond fund: R-squared ~57% daily, ~85% monthly (illiquidity gap). No alpha found, questioning its 58bp expense ratio.
- 12 quantitative mutual funds: comparable out-of-sample MSE to FF models. Monthly R-squared significantly higher than daily, confirming illiquidity effects.
- Sequential shrinkage produces the most stable betas over time, with smaller absolute coefficient changes than FF models for size and value factors.
Advantages over academic factor models
FF5’s factors are long-short portfolios sorted on firm characteristics, which are:
- Expensive to trade (tens to hundreds of bp per Novy-Marx & Velikov 2016)
- Subject to quiet data revisions
- Limited to equity-only characteristics
- Unstable: HML and CMA have a surprising 0.6 correlation, causing beta instability
The liquid model uses actual tradeable instruments and adds cross-asset factors (yield curve, credit, volatility, inflation).
Advantages over holdings-based models
Barra-style models:
- Factor returns change when the estimation universe changes
- Sweep all uncertainty into inferred returns (making significance testing impossible at asset level)
- Cannot capture cross-asset exposures (e.g., equity sensitivity to yield curves)
- Can be misled by aberrant holdings snapshots
The liquid model is returns-based and avoids all of these issues.
Primary use case: hedging overlays
The model is especially suited for large portfolios (pensions, endowments) that require derivative overlay solutions rather than portfolio restructuring. Because all factors are directly tradeable, hedging is cheap and actionable. Alpha estimates are more trustworthy because apparent alpha cannot hide behind illiquid or expensive-to-trade factor proxies.
Portability
The framework extends to other developed markets: EUR, JPY, AUD yield curves; European/Japanese equity indices; VSTOXX; European credit indices.