A factor model explains stock returns as the sum of exposures to a small number of common drivers, called factors. Instead of modeling each stock individually, researchers identify systematic characteristics (like company size, valuation, or price momentum) that predict differences in average returns across large groups of stocks. The construction process, from defining the characteristic to building a tradeable portfolio that isolates it, follows a specific methodology that has become the standard in academic finance and quantitative investing.

Understanding how factors are constructed matters because the methodology directly affects what the factor measures and whether the measured return premium is real or an artifact of the construction process. The same economic idea (for example, "cheap stocks outperform expensive stocks") can produce very different results depending on how "cheap" is defined, how the portfolio is formed, and how returns are calculated.

Conceptual Framework

The standard approach to factor construction was established by Eugene Fama and Kenneth French in their 1993 paper introducing the three-factor model. Their methodology, with variations, has been adopted by nearly all subsequent factor research. The process has three stages: define the characteristic, sort stocks into portfolios based on that characteristic, and compute the factor return as the difference between the high-characteristic and low-characteristic portfolios.

Defining the Characteristic

A factor characteristic is a measurable attribute of a stock that is hypothesized to predict future returns. Common examples include:

• Value: Ratios that compare a company's market price to its fundamental value, such as book-to-market ratio (book value of equity divided by market capitalization), earnings-to-price ratio, or cash-flow-to-price ratio.
• Size: Market capitalization (share price multiplied by shares outstanding). Smaller companies have historically earned higher average returns than larger companies.
• Momentum: Past return over a defined lookback period, typically the prior 2 to 12 months (excluding the most recent month to avoid short-term reversal effects). Stocks with higher past returns have historically continued to outperform.
• Quality/Profitability: Measures of fundamental business quality such as gross profitability (revenue minus cost of goods sold, divided by assets), return on equity, or earnings stability.
• Low Volatility: Historical return volatility or beta (sensitivity to market movements). Contrary to the prediction of the Capital Asset Pricing Model (CAPM), lower-risk stocks have historically earned similar or higher returns than higher-risk stocks.

The choice of characteristic involves judgment calls. Using book-to-market versus earnings-to-price as a value measure produces different stock rankings and different factor returns. Using 12-month versus 6-month lookback windows for momentum changes which stocks are classified as winners and losers. These choices are consequential, and researchers document their specific definitions so that results can be replicated and compared.

The Sorting Procedure

Once the characteristic is defined, stocks are sorted into portfolios based on their characteristic values. The standard Fama-French approach works as follows:

At the end of each formation period (typically June 30 for annual sorts, or month-end for monthly sorts), all eligible stocks are ranked by the characteristic. The ranked list is then divided into groups. The simplest division is into two groups: the top half and the bottom half. More common is a division into quantiles. Terciles (three groups) split stocks into the top third, middle third, and bottom third. Quintiles (five groups) and deciles (ten groups) provide finer granularity.

The Fama-French methodology uses NYSE breakpoints for the sort. This means that only New York Stock Exchange (NYSE) stocks are used to determine the cutoff values for each group, even though the resulting portfolios include stocks from all exchanges (NYSE, AMEX, and NASDAQ). This prevents the large number of small NASDAQ stocks from dominating the sort. Without this adjustment, the "small" portfolio would contain an overwhelming proportion of micro-cap stocks that are difficult to trade in practice.

Portfolio weights within each group are typically either equal-weighted (each stock gets the same weight) or value-weighted (each stock's weight is proportional to its market capitalization). Value-weighting is more common in academic research because it better represents the investable opportunity set and reduces the influence of tiny, illiquid stocks. Equal-weighting gives more influence to small stocks and tends to produce larger factor premiums, which can be misleading if those premiums are difficult to capture in practice.

Long-Short Portfolio Construction

The factor return is computed as the difference between the return of the high-characteristic portfolio and the return of the low-characteristic portfolio. This is called a long-short portfolio (or zero-cost portfolio, because it is self-financing in theory). The investor goes "long" (buys) the stocks with the desirable characteristic and goes "short" (sells) the stocks without it.

For the Fama-French value factor (HML, "High Minus Low"), the construction is: buy stocks with high book-to-market ratios (value stocks) and sell stocks with low book-to-market ratios (growth stocks). The return of this long-short portfolio is the value factor return for that period. If value stocks returned 8% and growth stocks returned 5%, the HML factor return is 3%.

The long-short construction isolates the return attributable to the characteristic itself by removing the market return. Both the long and short sides are exposed to the overall market, so the market component cancels out. What remains is the "pure" factor return: the return difference between stocks with high and low values of the characteristic, holding market exposure roughly constant.

Many factor constructions also control for other known factors. The Fama-French approach constructs HML by first sorting stocks into size groups (small and big), then sorting each size group by book-to-market. The HML factor is then the average of the value-minus-growth return across both size groups. This double-sort procedure ensures that the value factor is not contaminated by the size effect.

Multi-Factor Models

Individual factors are combined into multi-factor models that explain a larger portion of the cross-section of stock returns. The progression of standard models illustrates how the field has evolved:

• CAPM (1964): One factor: the market return. A stock's expected return depends only on its market beta (sensitivity to the overall market).
• Fama-French Three-Factor (1993): Market, size (SMB, "Small Minus Big"), and value (HML). Added two factors that the CAPM could not explain.
• Carhart Four-Factor (1997): Added momentum (UMD, "Up Minus Down" or WML, "Winners Minus Losers") to the Fama-French three factors.
• Fama-French Five-Factor (2015): Added profitability (RMW, "Robust Minus Weak") and investment (CMA, "Conservative Minus Aggressive") to the original three factors.
• Q-Factor Model (Hou, Xue, Zhang, 2015): An alternative specification using market, size, investment, and profitability factors, motivated by investment-based asset pricing theory.

Each model is used as a benchmark for evaluating investment strategies. If a strategy's returns can be fully explained by its factor exposures (loadings on the known factors), then the strategy is not generating "alpha" (excess returns beyond what the factors predict). The strategy is simply packaging known factor tilts. If the strategy generates returns that the factor model cannot explain, those unexplained returns are alpha, and they suggest the strategy has identified a genuine source of return beyond the standard factors.

Risk Architecture

The Factor Zoo Problem

Academic researchers have published hundreds of factors that purportedly predict stock returns. Harvey, Liu, and Zhu (2016) catalogued over 300 published factors and noted that most would fail more stringent statistical tests. The proliferation of factors, sometimes called the "factor zoo," raises fundamental questions about how many of these are genuine return drivers versus statistical artifacts of data mining.

The problem stems from the publication process. Researchers test many potential characteristics against the same historical return data. Those that show statistically significant results get published; those that do not are discarded. This selection bias means that the published literature overrepresents factors that happened to work in-sample, including some that were merely lucky. Standard statistical significance thresholds (t-statistic above 2.0) are insufficient when hundreds of tests have been conducted on the same data. Harvey, Liu, and Zhu suggest a threshold of approximately 3.0 to account for the multiple testing problem.

Sensitivity to Construction Choices

Factor premiums are sensitive to seemingly minor methodological decisions. Research has shown that the following choices can materially affect the measured premium:

• Characteristic definition: Using book-to-market versus earnings-to-price for value produces different factor returns. Novy-Marx (2013) showed that gross profitability and operating profitability generate substantially different factor premiums.
• Weighting scheme: Equal-weighted factor portfolios typically show larger premiums than value-weighted portfolios. The difference reflects the contribution of small, illiquid stocks that are difficult to trade at the prices assumed in the backtest.
• Rebalancing frequency: Annual rebalancing (standard for value and size) versus monthly rebalancing (common for momentum) affects turnover, transaction costs, and the measured premium.
• Universe definition: Including or excluding micro-cap stocks, international markets, or specific sectors changes the measured factor premium. Factors that appear strong in a broad universe may weaken when the universe is restricted to large, liquid stocks.

Known Limitations

Practical Considerations

Evaluating Factor Claims

When evaluating whether a proposed factor represents a genuine return driver, several criteria help separate robust findings from data-mined artifacts:

• Economic rationale: Is there a plausible reason why this characteristic should predict returns? Factors can be justified either as compensation for risk (investors demand higher returns for bearing the risk associated with the characteristic) or as a behavioral anomaly (systematic mispricing caused by investor biases). Factors without a clear economic explanation are more likely to be statistical artifacts.
• Statistical robustness: Does the factor survive stringent statistical tests? A t-statistic above 3.0 (rather than the traditional 2.0) provides a more reliable threshold given the multiple testing problem. The premium should be present across different sample periods and not driven by a few extreme observations.
• Out-of-sample evidence: Does the factor work in time periods, countries, and asset classes beyond the original study? A value premium documented in U.S. equities from 1960 to 1990 is more credible if it also appears in European and Asian markets and in post-publication periods.
• Implementability: Can the factor be captured in a real portfolio after transaction costs? Factors that rely on frequent trading of illiquid small-cap stocks may generate paper returns that cannot be realized in practice.

From Academic Factor to Investment Product

Translating an academic factor into a real investment portfolio involves several practical modifications. Academic factors are rebalanced at fixed intervals (monthly or annually), use the full universe of stocks including illiquid micro-caps, and ignore transaction costs. Real implementations must address each of these:

• Universe restrictions: Excluding stocks below a minimum market capitalization or trading volume threshold. This reduces the measured premium but ensures the portfolio can be executed.
• Turnover management: Using buffer zones around sort breakpoints so that small changes in a stock's characteristic do not trigger trades. A stock at the 29th percentile that moves to the 31st percentile should not necessarily change portfolios.
• Transaction cost budgeting: Limiting turnover to a level where the net-of-cost premium remains positive. High-turnover factors like short-term momentum require careful cost management.
• Multi-factor integration: Combining multiple factor exposures in a single portfolio to capture diversified factor premiums while minimizing turnover. This is more efficient than holding separate single-factor portfolios and rebalancing each independently.

Related Concepts and Models

Further Reading

• Fama, E.F. and French, K.R. (1993). "Common Risk Factors in the Returns on Stocks and Bonds." Journal of Financial Economics, 33(1), 3–56.
• Harvey, C.R., Liu, Y., and Zhu, H. (2016). "...and the Cross-Section of Expected Returns." The Review of Financial Studies, 29(1), 5–68.
• Fama, E.F. and French, K.R. (2015). "A Five-Factor Asset Pricing Model." Journal of Financial Economics, 116(1), 1–22.
• McLean, R.D. and Pontiff, J. (2016). "Does Academic Research Destroy Stock Return Predictability?" The Journal of Finance, 71(1), 5–32.
• Novy-Marx, R. (2013). "The Other Side of Value: The Gross Profitability Premium." Journal of Financial Economics, 108(1), 1–28.