The Economic Rationale Behind a Signal
A pattern found by mining historical data is not the same thing as an edge — a signal only deserves real capital once there's a plausible story for why the world should keep producing that pattern going forward.
Prerequisites: Where Alpha Ideas Come From, A Taxonomy of Alpha Sources
Backtest a large enough set of historical patterns and some of them will look profitable purely by chance — with enough attempts, randomness alone produces winners. That is why a signal's backtest performance alone is not evidence of a real edge. What separates a durable signal from a lucky accident is whether there's a plausible reason the pattern should keep showing up: some economic mechanism, some group of market participants who are structurally forced to keep generating it.
Researchers generally sort rationales into three buckets. Risk compensation: the signal earns money because it bears a risk other investors won't hold (small, illiquid stocks tend to earn more because they're harder to exit in a crisis). Behavioral bias: the signal exploits a systematic error other participants keep making (momentum partly reflects investors underreacting to news and adjusting slowly). Structural or informational edge: the signal captures a flow that has to happen regardless of price (index funds must buy an added stock on the rebalance date, or an analyst's private channel checks give an early read on earnings).
A signal without a plausible economic story is a coin flip that happened to come up heads in the data you tested — the story is what gives you a reason to expect the pattern to persist out of sample, rather than just describing what already happened.
Story first, or number first?
Worked example
Two researchers each present a signal with an identical 5-year backtested Sharpe ratio of 1.1. Signal A: stocks that get added to a major index tend to rise in the days before their scheduled inclusion date. The story: index funds tracking that benchmark are contractually required to buy the stock on inclusion, and some price-insensitive buying pressure is predictable well in advance — a mechanism that doesn't depend on any information about the company at all, only on the mechanics of index tracking.
Signal B: stocks whose ticker symbol contains the letter "Q" have historically outperformed. There is no economic mechanism connecting a letter in a ticker to future returns; the pattern, however statistically significant it tests, is almost certainly a product of having tested enough arbitrary groupings that one of them looked good by chance. A researcher should trust Signal A's backtest far more than Signal B's, even with matching statistics, because only one of them has a reason to keep working next year.
Do not be reassured by a story invented after a data-mined pattern is found — a plausible-sounding explanation can be constructed for almost any pattern after the fact. The rationale should ideally predate the data search, or at minimum should make a testable prediction about where else the same mechanism should show up, which you can then go check.
Practice in interviews
Further reading
- Grinold and Kahn, Active Portfolio Management (ch. 1)