How Would You Test Whether Momentum Is a Real Alpha Signal in Stocks?

This is a methodology question in a quantitative-research data-analysis interview. You will not write code; the interviewer expects you to think like a statistician and will interrupt with sharp follow-ups that have objectively correct or incorrect answers, covering standard errors, biases, and test validity.

Suppose you hold the view that stocks with strong trailing one-year performance tend to continue outperforming in the near term — a momentum effect. Imagine you are given a long daily price history for a broad equity universe. How would you rigorously test whether momentum is a genuine alpha signal — that is, whether it has statistically and economically meaningful predictive power for future returns beyond what known risk factors explain?

Lay out the complete methodology: how you would define the signal, set up the tests, select the test statistics, protect against biases, and judge whether the alpha would hold up in practice. Use classical statistics and linear models rather than black-box machine learning.

Frame it as a hypothesis test. State a null hypothesis that the signal has no predictive power for future cross-sectional returns. Then choose statistics that capture the signal-to-future-return relationship: the period-by-period information coefficient (IC), the average return of a long-short decile portfolio, or the slope from a cross-sectional regression. Each yields a time series that can be tested with a t-test.

Standard errors are a trap. With a 12-month formation window and monthly rebalancing, consecutive observations overlap substantially, so the test-statistic series is autocorrelated and naive t-statistics overstate significance. Consider HAC (Newey–West) corrections — and also the multiple-testing issue if many signal variants were tried before this one.

Alpha vs. risk premium. A profitable signal is not alpha if it simply loads on known factors. Run a regression of the strategy's return series on standard factor returns (a linear model) and test whether the intercept is significantly positive.

Constraints & Assumptions

• Assume a broad, liquid equity universe of a few thousand names, with at least 10 years of daily point-in-time prices adjusted for splits and dividends, and with delisted stocks included.
• You may define "momentum" precisely; a common convention is the trailing 12-month return excluding the most recent month, but you should justify your choice.
• No code is required; the interviewer wants the methodology, the statistics, and the reasoning behind each choice.
• Stay within classical statistics and linear models; complex machine-learning models are outside the scope of this round.

Clarifying Questions to Ask

• What universe and region are we testing — large-cap US, global, or something else — and over what sample period?
• At what horizon should the signal predict returns — days, one month, one quarter — and how often would the strategy rebalance?
• Is the bar "statistical predictability" or "implementable alpha" — should the test be gross or net of transaction costs?
• Which known risk factors am I expected to control for — market, size, value, and so on?
• Is the momentum definition fixed, or am I free to specify the formation window, skip period, and cross-sectional versus time-series construction?
• Do we care about long-short performance or long-only excess return relative to a benchmark?

What a Strong Answer Covers

Follow-up Questions

• Your long-short momentum portfolio has a full-sample t-statistic of about 2.2. Do you conclude the alpha is real? What evidence would change your mind?
• You chose a 12-month formation window with monthly rebalancing, so adjacent observations overlap for 11 months of formation data. How exactly does that bias inference, and what do you do about it?
• Once standard factors are added to the regression, the intercept shrinks and becomes insignificant. What are the possible interpretations, and what would you test next?
• How would your methodology change if the signal is to be traded at meaningful size, and what additional analyses matter before committing real capital?

Overview: This question tests statistical hypothesis testing, econometric skills in time-series and cross-sectional return analysis, factor-regression proficiency, and the ability to recognize biases such as autocorrelation and multiple testing when judging whether a trading signal represents true alpha.