Factor Construction Methodology

A factor is not a thing found in nature; it is a set of choices about how to define, measure, and execute a trading signal. Different methodologies applied to the same underlying idea can produce startlingly different results.

The construction problem

Suppose a researcher observes that companies with low ratios of market price to book value historically outperform the market. That observation is not a factor yet—it is a puzzle. To turn it into a tradeable factor, the researcher must make dozens of concrete decisions:

• How is “low book value relative to price” defined? Price-to-book? Price-to-equity? Adjusted for intangible assets?
• Over what period is the book value measured? The most recent quarter? A trailing twelve-month average?
• What is “low”? The bottom quintile? The bottom 10%? Stocks below a certain absolute threshold?
• Once ranked by this signal, how are the stocks weighted in the portfolio? Equal-weight? Market-cap weight? Risk-parity weight?
• How often is the portfolio rebalanced? Monthly, quarterly, annually?
• What transaction costs and slippage are assumed in the backtest?

The reason these choices matter is not pedantic: they can double or halve measured returns. Most researchers discovering a factor employ the methodology that produces the best historical results, then report that result as “the factor.” Different choices yield different factors.

Signal definition

The definition of the signal is the core choice. A value factor might use:

• Price-to-earnings ratio (plain or trailing earnings)
• Price-to-book ratio (tangible book value or adjusted for intangibles)
• Price-to-sales ratio (revenue-based, less subject to earnings manipulation)
• A composite of several metrics
• Book value plus dividend yield (yield-adjusted value)

Each definition identifies a partially overlapping but distinct set of “cheap” stocks. Backtests using price-to-book might show a 6% annualised outperformance; the same signal using price-to-sales might show 4%. Neither is “right”—they measure different things. The researcher chooses the definition that performs best historically, introducing survivorship bias and overfitting risk.

The choice also determines which stocks are selected. A value signal on price-to-book might identify a portfolio of industrials and financials; the same signal on price-to-sales might identify retailers and consumer goods. These are materially different portfolios with different risks.

Portfolio formation and ranking

Once a signal is defined, the researcher must decide how to form the portfolio:

Decile or quintile breakpoints. The researcher ranks all stocks by the signal and divides them into ten (or five) equal-weight buckets. The lowest decile—the “cheapest” 10%—becomes the long portfolio; the highest decile becomes the short. This is transparent and easy to implement, but it creates a hard cliff: a stock just barely inside the cutoff behaves identically to one far inside it.

Continuous weighting. Instead of breakpoints, the researcher weights each stock by its signal rank. Cheaper stocks get larger positions. This is smoother but requires defining the weighting function (linear, square-root, log?), which is another choice.

Quantile tilts. Rather than creating a hard long-short portfolio, the researcher tilts a market-cap-weighted index to overweight cheap stocks and underweight expensive ones. This is closer to a real long-only portfolio but dilutes the factor signal’s purity.

Each approach produces a different return profile, turnover, and risk distribution.

Weighting scheme

The portfolio is formed; now, how are the positions sized?

Equal-weight. Each stock gets the same dollar amount. This is simple but creates a drag on large-cap stocks (expensive to trade) and sometimes generates huge idiosyncratic bets on tiny illiquid names.

Market-cap weight. Positions are sized by the stock’s market cap within its decile. This mimics typical index construction and reduces liquidity drag but concentrates risk into the largest names in the decile, which may not be the ones with the strongest signal.

Risk-parity weight. Positions are sized to equalise volatility across stocks. A highly volatile cheap stock gets a smaller position; a stable one gets a larger position. This reduces idiosyncratic risk but requires estimating future volatility, which is notoriously unstable.

Dollar-volume weight. Positions are sized by the average daily trading volume of the stock, ensuring the portfolio can exit without moving the market too far. This reduces liquidity risk but biases the portfolio toward liquid names, many of which may not be cheap.

A momentum factor using equal-weight deciles produces returns that look very different from the same factor using market-cap weighting.

Rebalancing frequency

How often the portfolio is reset to its target composition also shapes returns:

Monthly rebalancing. The portfolio is reconstituted at the end of each month. This is frequent enough to capture the signal fresh but generates high turnover and high transaction costs, especially in small-cap or illiquid segments.

Quarterly rebalancing. A middle ground. The signal is refreshed, but not so often that the bid-ask spread becomes the dominant driver of returns. Academic papers often use quarterly to match earnings announcement cycles.

Annual rebalancing. Lower turnover, lower costs, but the portfolio drifts. A stock that was cheap in January may have doubled by December and is no longer cheap, yet it is still held.

The choice is a tradeoff between signal freshness and transaction costs. Monthly rebalancing shows higher volatility and higher costs; annual rebalancing shows lower costs and often better risk-adjusted returns.

Survivorship and look-ahead bias

Construction methodology is where many academic factor papers hide overfitting. A careful analysis might try 50 different signal definitions, 10 weighting schemes, and 4 rebalancing frequencies. After 2,000 backtests, the researcher selects the one with the highest Sharpe ratio. That result is the product of data mining, not a true discovery.

Worse, the backtest often includes data that would not have been available at the time of decision. A signal defined as price-to-book-value might use the most recent quarter’s balance sheet, even if it is published only weeks later. A backtest that buys stocks at month-end using data that becomes public at month-end is employing look-ahead bias.

Real implementations must also decide: Do we buy at the close? At market-on-open? Do we use the average price during the day? All choices that have been optimised away in the backtest.

How methodology affects practical deployment

A factor strategy backtested with annual rebalancing, equal-weight deciles, and frictionless execution might show a 10% annualised excess return with a 12% annualised volatility. The same factor, deployed in practice with monthly rebalancing, market-cap weighting to reduce micro-cap exposure, and realistic transaction costs, might deliver 4–5% annualised excess return with 11% volatility. The difference is not fraud—it is methodology.

This is why factor funds often publish two types of results: backtested (to show what they are trying to capture) and live (net of fees, trading costs, and real market impact). The backtested result attracts investors; the live result is what they actually earn.

The path forward

Professional factor researchers understand that construction methodology is not a detail to fix and forget. It is a source of risk and opportunity. Some teams deliberately choose methodologies that are robust to small changes—annual rebalancing, market-cap weighting, broad universes—because they distrust their own backtests. Others optimise methodology for operational simplicity or risk control, even if it costs a few basis points of alpha.

The key is to be explicit: publish the rules, stick to them, and separate the signal’s true efficacy from the noise introduced by the choices used to measure it.

See also