Factor strategies: ranking, neutralization, and portfolio formation

A factor is a repeatable way to compare assets

In strategy research, a factor begins with a characteristic or signal that can be measured across many assets. Value compares price with an economic anchor. Momentum compares recent performance. Quality compares features such as profitability or balance-sheet strength. Low-volatility compares measured risk.

The measurement is not yet a portfolio. To become investable, the researcher must decide which assets are eligible, when their data was known, how scores are made comparable, which exposures are removed, how positions are weighted, and how often the portfolio changes. Influential factor conventions make those sorting and long-short portfolio decisions explicit rather than treating the score itself as investable.

Two funds can both say they trade “value” while holding very different portfolios. One may buy the cheapest fifth of the whole market. Another may rank within each sector, exclude illiquid stocks, combine several valuation ratios, and cap every position. The label is only the beginning of the design.

Factor strategies answer a relative question

Most equity factors are cross-sectional: at one decision time, they compare each asset with its peers. A score of 0.8 may mean “higher than 80% of the eligible universe,” not that the asset has positive expected return in absolute terms.

That distinction matters during broad declines. A top-ranked stock can fall while still outperforming the bottom-ranked stocks. A long-only factor portfolio mixes the relative signal with general market exposure; a long-short spread tries to isolate more of the relative effect but introduces shorting, financing, and capacity assumptions.

Some cross-asset signals, such as carry, use different economics. Carry asks what return is implied by holding an asset or rolling a contract if relevant market conditions do not move against the position. Futures roll yield, currency forward differentials, and bond yield relationships are not interchangeable, but they share the factor workflow: define a comparable measure, rank a point-in-time universe, and form a portfolio.

Why factor returns might exist

There are several competing explanations:

• Risk compensation: the portfolio may perform on average because it loses in states investors particularly dislike.
• Behavior: investors may underreact, extrapolate, prefer glamorous assets, or neglect less exciting ones.
• Constraints: leverage limits, benchmarks, mandates, taxes, and shorting restrictions can prevent prices from adjusting fully.
• Institutional flows: slow rebalancing and crowded portfolio conventions can create persistent relative demand.

These explanations predict different failures. A risk premium may remain available but suffer severe losses in its insured state. A behavioral effect may decay after adoption. A constraint-based effect may change when financing or market structure changes.

Backtest returns alone do not identify the mechanism.

The portfolio-construction pipeline

Every step should be reproducible from information available at the rebalance.

Ranking and neutralization in plain language

Suppose ten eligible companies receive a profitability score. Ranking orders them from strongest to weakest. A simple portfolio might equally weight the top three and avoid the rest.

Now suppose all three top companies are banks. The result combines profitability with a concentrated financial-sector position. Sector neutralization would rank banks against banks, industrial firms against industrial firms, and so on, then select within each group. The portfolio expresses relative profitability inside sectors but may weaken or eliminate a genuine cross-sector effect.

Neutralization is not automatically better. It states which risks the strategy is not allowed to use. The unneutralized and neutralized results answer different questions and should be shown separately.

A complete illustrative rebalance

At the final trading day of March:

1. Freeze the point-in-time universe using eligibility and liquidity rules known that day.
2. Use only accounting statements published by the decision cutoff. A December fiscal-year figure filed in February may be eligible; a later restatement is not substituted into history.
3. Transform the signal according to a declared rule and rank within sectors.
4. Select the top and bottom quantiles, apply position and sector caps, and compute target weights.
5. Execute at the next modeled opportunity and hold until the next scheduled rebalance.
6. Record every entry, exit, rank migration, corporate action, borrow restriction, and transaction cost.

If an asset crosses the quantile cutoff by one rank, the entire target position can change. Continuous score weights or buffered membership can reduce that cliff, but they create new portfolio rules.

How to research a factor honestly

Start with a transparent single signal and equal-weight quantiles. Show:

• long-only, short-only, and long-short results where feasible;
• sector, country, size, beta, and other factor exposures;
• turnover, spread, impact, and borrow costs;
• point-in-time versus latest-data differences;
• performance by decade and market state;
• concentration and capacity; and
• sensitivity to reasonable peer groups and rebalance schedules.

Then add neutralization, risk weighting, or signal combinations one at a time. Compare every extension with the unchanged baseline. A more complex version must improve a defined research outcome, not merely the historical Sharpe ratio.

Use later dates or markets as genuine evaluation. Factor definitions can be selected from hundreds of characteristics, so the search and publication process creates a large multiple-testing burden.

What can invalidate the result

• Point-in-time data removes the apparent ranking.
• Microcaps or impossible short positions produce most of the spread.
• One sector, country, or market-beta exposure explains the return.
• Transaction costs rise faster than the signal when rebalancing is made more frequent.
• The effect depends on one arbitrary treatment of negative or missing values.
• Many correlated definitions were tried and only the best was reported.
• Crowded factor positions unwind together despite different names.

A factor strategy is credible only when its data lineage, portfolio formation, and implementation are as clear as its economic story.

Further reading

• Fama and French, “Common Risk Factors in the Returns on Stocks and Bonds” (1993) — Provides an influential, explicit example of turning characteristics into sorted portfolios and long-short research factors. It is a construction convention, not proof that every product using the same label has the same exposure.
• Asness, Moskowitz, and Pedersen, “Value and Momentum Everywhere” (2013) — Applies comparable value and momentum constructions across several asset classes, showing both commonality and the importance of market-specific definitions.
• Koijen, Moskowitz, Pedersen, and Vrugt, “Carry” (2018) — Defines carry across global asset classes and makes clear that currency, bond, commodity, equity-index, and option carry require different underlying measurements.
• Harvey, Liu, and Zhu, “… and the Cross-Section of Expected Returns” (2016) — Quantifies the multiple-testing problem created by a large and growing factor literature and argues for a higher evidentiary hurdle than an isolated conventional test.
• Novy-Marx and Velikov, “A Taxonomy of Anomalies and Their Trading Costs” (2016) — Estimates how turnover and cost-mitigation rules alter anomaly returns, illustrating why a paper rank is not yet an investable factor.
• Hou, Xue, and Zhang, “Replicating Anomalies” (2020) — Rebuilds 452 published anomalies under common procedures; many weaken after microcap, weighting, and multiple-testing choices, making construction transparency central to interpretation.