Spanning Tests for a New Factor

Suppose someone hands you a new trading factor and its backtest looks great. Before adding it to a portfolio that already runs value, momentum, and quality, the obvious question is: does this new factor do anything those three don't already do? If its returns can be built out of a combination of the existing three, it isn't a new source of return — it's the same return, relabeled. A spanning test answers exactly that question with a regression.

Setting up the regression

The mechanics are a single time-series regression. Take the new factor's monthly returns as the dependent variable, and the returns of the existing factors (say, market, value, momentum, quality) as the independent variables:

$$R_{new,t} = \alpha + \beta_1 R_{value,t} + \beta_2 R_{momentum,t} + \beta_3 R_{quality,t} + \varepsilon_t$$

In words: the new factor's return each month is explained by how much it moves with value, momentum, and quality, plus a leftover intercept $\alpha$. If $\alpha$ is close to zero and not statistically significant, the new factor earns nothing that a combination of the old three didn't already deliver. If $\alpha$ is large and significant, the new factor has genuine, unspanned return — something worth adding.

Worked example

A researcher proposes a "management quality" factor earning 0.8% per month on average. Regressed against value, momentum, and quality, it loads at $\beta_{quality} = 0.7$ and small betas elsewhere, with an intercept of $\alpha = 0.05\%$ per month and a t-stat of 0.4. Conclusion: the new factor's return is almost entirely explained by its overlap with the existing quality factor — it is spanned, and adding it to a book that already holds quality contributes essentially nothing new.

Contrast that with a second candidate, a "supply-chain disruption" factor earning 0.6% per month, which loads weakly on all three existing factors ($\beta$'s under 0.15) and leaves an intercept of $\alpha = 0.45\%$ per month with a t-stat of 2.6. This one is not spanned — it survives the test and is a genuine candidate for inclusion.

What this means in practice

Spanning tests are the standard first filter a quant researcher applies before adding a new signal to a live model, and they're what separates "this backtested well" from "this is actually additive." A factor can have a great Sharpe ratio on its own and still fail a spanning test if it happens to correlate strongly with something already in the book.