Combine factor scores without hiding missing data

A composite needs a measurement contract

A composite score combines descriptors measured on different scales. Earnings yield might be a percentage, momentum a cumulative return, and leverage a ratio for which lower values are preferred. Averaging raw values makes the largest numerical scale dominate. Define the direction and transformation of each component before combining them, and preserve the raw inputs so a reviewer can reconstruct the ranking.

MSCI's May 2022 Quality Indexes methodology illustrates winsorization followed by standardized scores. It is cited here as a historical methodological example, not as the current specification of every quality index. An independently designed equity score must declare its own clipping thresholds and standard-deviation convention. Borrowing the word standardized does not establish that two implementations produce equivalent exposures.

Equal composite values can conceal different businesses

Suppose all component scores are already oriented so higher means preferred. Fictional Stock L has value, profitability and momentum scores of 1.2, -0.4 and 0.7. Stock M has 0.3, 0.8 and 0.4. Both average 0.5. L's strong value score compensates for weak profitability, whereas M is more balanced. The identical composite does not mean identical risks.

If L's profitability input becomes unavailable, averaging its remaining scores would produce 0.95. That apparent improvement is caused entirely by missing data. A complete-case policy would instead exclude L; a neutral-imputation policy would produce about 0.63 and should carry a missingness flag. Neither choice is universally correct. The rule must be fixed before measuring the strategy and its coverage.

Separate date-local scaling from fitted transformations

Ranking the observable cross-section on each formation date is different from fitting a transformation across the entire historical dataset. A contemporaneous percentile can use all eligible stocks known at that time. A learned mapping from features to expected returns, or a historical scaling parameter, must be estimated from the allowed training sample only. Future observations cannot influence either merely because preprocessing happens before model fitting.

State how ties, constant-valued components and extreme ratios are handled. A zero cross-sectional standard deviation makes a z-score undefined; replacing it with a tiny denominator can create meaningless precision. A deterministic neutral score or explicit component exclusion may be preferable if declared. Keep issuer identity as a stable tie-breaker instead of whichever row happens to appear first in a file.

Inspect the composite before judging performance

Measure rank correlations and component contributions. Three highly correlated value proxies do not create three independent sources of information, and equal component weights need not produce equal risk contributions. Inspect the top and bottom selections, their sector composition and their missing-data rates. A score that works only where reporting is unusually complete may primarily select a size or coverage characteristic.

Keep a small, predeclared set of alternatives and record them as separate trials. Do not tune clipping, weights and missing-data rules repeatedly on the same final period. The example establishes how composite arithmetic can conceal a changed information set; it does not identify the best aggregation method. A useful result must survive its documented universe, timing, cost and validation assumptions.

Inspect the example

Sources & reproduction

Primary references support the definitions and method. Numerical examples are illustrative unless a source record is explicitly identified.

• MSCI Quality Indexes Methodology, May 2022 (historical version)

Continue the review