Factor Timing
Valentin Haddad et al. (3 authors), 2020 · The Review of Financial Studies
What the paper asks for.
Time-varying expected returns of the market and five anomaly PCs are predictable from factor-specific book-to-market signals, so an unconstrained mean-variance factor-timing portfolio that sets monthly weights to Sigma^{-1} E_t[r_{t+1}] improves on static factor investing.
- Type
- fundamental
- Implementation
- pandas
- Rebalance
- monthly
- Long-short
- yes
- Data
- Bundled public task data
- Operational scope
- Use the authors' 55-anomaly release, an aggregate market book-to-market proxy, and a mild ridge allowance for predictive regressions.
- 01 Data loading
- 02 Feature engineering
- 03 Model fitting
- 04 Signal generation
- 05 Portfolio construction
- 06 Execution
19 properties · 80 attempts.
Each row is one requirement the strategy must satisfy. The bar reports the share of observed checker verdicts that pass; missing checker output is not converted into a property verdict.
- A1no backfill75/77
- A2no global normalize60/77
- A3no future data access64/77
- A4causal consistency65/77
- A6train test temporal order64/77
- A8transaction timing74/77
- A9signal delay test77/77
- A10e2e metric consistency0/77
- B1anomaly decile count§Sec 3.1 “sort them into 10 value-weighted portfolios for each of the 50 characteristics”provenance · paper-explicitstage · feature eng77/77
- B2long short 10 minus 1§Sec 3.1 “return on portfolio 10 minus the return on portfolio 1”provenance · paper-explicitstage · feature eng77/77
- B3market adjustment beta in sample only§Sec 3.1 “betas ... are estimated using only the first half”provenance · paper-explicitstage · feature eng60/77
- B4pca calibration window§Sec 3.2 “we estimate Q and Lambda using only the first half of the data”provenance · paper-explicitstage · model fit64/77
- B5num pcs equals 5§Sec 3.2 “these five components jointly explain nearly two-thirds of the total variation in returns”provenance · paper-explicitstage · model fit77/77
- B6factor set is market plus 5 pcs§Sec 3.2 “Z_{t+1} = (R_{mkt,t+1}, PC_{1,t+1} ... PC_{5,t+1})”provenance · paper-explicitstage · model fit73/77
- B7univariate predictive regression own bm§Sec 3.3 “we construct a single predictor for each portfolio: we use its net book-to-market ratio”provenance · paper-explicitstage · model fit73/77
- B8unconstrained mean variance weights§Eq. 10 “optimal portfolio weights are w_t = Sigma^{-1} E_t[Z_{t+1}]”provenance · paper-explicitstage · portfolio77/77
- B9monthly rebalance frequency§Sec 3.1 “monthly returns from January 1974 to December 2017”provenance · paper-explicitstage · execution77/77
- B10forecast formula affine§Sec 3.3 “standard predictive regressions on valuation ratios”provenance · paper-explicitstage · signal gen77/77
- B11market adjustment applied before pca§Sec 3.1 & 3.2 “we market-adjust ... We construct PCs from the 50 anomaly portfolios”provenance · paper-explicitstage · feature eng76/77
The lowest-passing observed requirement is e2e metric consistency: 0 of 77 observed checker verdicts pass. The full property list above shows whether errors concentrate in domain-level validity or the paper-specific strategy logic.