Python pipeline
End-to-end workflow from data preparation through estimation, forecasting, tables and figures.
Empirical Asset Pricing · Python · 2026
Macro-Regime Dependent
Asset Pricing
Testing whether macroeconomic state dependence in factor exposures and premia translates into better real-time forecasts of U.S. industry returns.
State dependence is detectable. Real-time forecastability is not.
Test Assets
49
value-weighted U.S. industries
Full Sample
683
monthly observations · Jul 1969–Jun 2026
OOS Evaluation
437
forecast months · Jan 1990–Jun 2026
Real-Time Forecasts
21,413
industry-month predictions
01
Research question
Does conditioning a standard factor model on a predetermined macroeconomic stress regime improve real-time forecasts of industry returns?
Test assets
49 value-weighted U.S. industry portfolios
Factor model
Fama-French five factors + momentum
Macro state
Predetermined real-time Sahm labor-market stress signal
Benchmark
Unconditional FF6
vs.
Test model
Regime-Conditional FF6
02
Research design
01
Data
49 industry portfolios and the six FF6 factors.
02
Macro State
Predetermined NORMAL / STRESS classification.
03
Estimation
Unconditional FF6 versus regime-conditional FF6.
04
Forecast
Expanding-window next-month industry forecasts.
05
Evaluation
Forecast errors, rank IC and portfolio illustration.
Every out-of-sample forecast is formed using parameters and information available before the return month.
03
Real-time information design
No look-ahead information enters the forecast.
Return month t is assigned the latest Sahm observation that would have been observable under the study's information-timing convention, using the reading from t−2.
t−2
Latest observable macro signal
Regime information available before forecast formation.
t−1
Estimate & forecast
Model parameters are estimated only through the prior month.
t
Realized industry return
The forecast is evaluated only after the return is observed.
Macro regime
Jul 1969–Jun 2026
Real-time Sahm stress signal
04
In-sample evidence
Factor relationships do change across macroeconomic states.
The conditional specification identifies statistically meaningful state dependence in a subset of industries, a pronounced change in the SMB premium, and lower in-sample pricing errors.
State-dependent exposures
11 / 49
industries remain significant after Benjamini-Hochberg FDR control at 5%.
Clearest premium shift
SMB
−0.109% per month in NORMAL versus +0.835% in STRESS.
Pooled pricing RMSE
0.393%
conditional FF6 versus 0.479% for unconditional FF6.
Cross-sectional evidence
FDR-significant industries
Factor returns
Premia by regime
In-sample fit
Pricing-error comparison
What this establishes
Conditioning improves the description of realized returns in sample. The harder question is whether those relationships remain useful when the regime and all model parameters must be estimated in real time.
05
Real-time forecasting test
State dependence does not translate into better real-time forecasts.
Across 21,413 industry-month predictions, the conditional FF6 model delivers slightly higher aggregate forecast errors than the unconditional benchmark and essentially no cross-sectional ranking ability.
Conditional RMSE
6.752%
real-time OOS forecast error.
Unconditional RMSE
6.742%
benchmark OOS forecast error.
OOS R²
−0.312%
relative to unconditional FF6.
Clark-West
p = 0.398
one-sided HAC test.
Forecast performance
Positive = conditional model ahead
Cumulative squared-error difference
Central finding
State dependence ≠ forecastability
In sample
- 11 of 49 industries show FDR-significant state-dependent FF6 exposures.
- SMB displays the clearest state-dependent factor premium.
- Conditional pricing errors are lower.
Real-time out of sample
- Aggregate forecast RMSE does not improve.
- OOS R² versus unconditional FF6 is negative.
- Cross-sectional ranking ability is negligible.
06
Economic illustration
Top-10 portfolios
10 bps trading costs
Cumulative wealth
Conditional Top-10 CAGR
12.14%
Unconditional Top-10 CAGR
10.90%
Market benchmark CAGR
11.05%
Conditional FF6 alpha
−1.55% p.a.
t = −1.16 · p = 0.247
Higher raw returns in the illustrative conditional Top-10 portfolio do not translate into statistically significant positive factor-adjusted abnormal performance.
07
Robustness
The core forecasting conclusion remains unchanged under alternative factor sets, estimation windows, test assets and regime specifications.
| Specification | OOS R² vs. benchmark | Clark-West p | Conclusion |
|---|---|---|---|
| FF5 | −0.266% | 0.384 | No improvement |
| 20-year rolling window | −0.660% | 0.510 | No improvement |
| 30 industry portfolios | −0.258% | 0.382 | No improvement |
| Continuous Sahm signal | −0.196% | 0.072 | Suggestive only |
CFNAI and NBER alternatives are treated as ex-post / in-sample robustness checks rather than real-time forecasting exercises.
08
Research implementation
Walk-forward testing
437 sequential forecast months evaluated with an expanding historical information set.
No-look-ahead design
Explicit timing controls ensure the regime and estimated parameters precede the forecast return.
Statistical inference
HAC inference, Wald tests and Benjamini-Hochberg false-discovery-rate control.
Robustness framework
Alternative factor sets, estimation windows, test assets and macro-state definitions.
Reproducible outputs
Research tables, diagnostics and website figures are generated programmatically from project outputs.
Full research
Explore the complete study.
The full Research Note contains the methodology, statistical evidence and robustness analysis. A shorter Executive Summary provides the research question, design and principal findings at a glance.
Data
49 value-weighted U.S. industry portfolios and factor returns from the Kenneth R. French Data Library; real-time Sahm labor-market stress data from Federal Reserve Economic Data.
Sample
Full sample: July 1969–June 2026. Real-time out-of-sample evaluation: January 1990–June 2026. The six-factor specification combines the Fama-French five factors with momentum.
Research scope
Independent empirical asset-pricing research project. Results on this page are condensed from the full Research Note and are presented for academic and portfolio purposes only, not as investment advice.