Where Alpha Ideas Come From

Ask a junior researcher where to find a new signal and the honest first answer is: not by staring at a price chart hoping a pattern jumps out. Productive quant research groups pull ideas from a short list of repeatable sources, and which source an idea comes from changes how you should test it — an idea grounded in a real economic mechanism deserves a different kind of scrutiny than one found by scanning ten thousand correlations for the highest one.

The four reliable sources

Economic mechanisms. Start from a story about why a price should move, then look for it in data. Momentum exists, plausibly, because investors underreact to news and then herd once the trend is visible; value exists, plausibly, because cheap stocks carry real distress risk investors demand compensation for. An idea sourced this way has a built-in falsification test: if the mechanism is real, the effect should show up where the mechanism should apply (large-cap value, small-cap value, international value) and weaken where it shouldn't.

Market structure and microstructure. These ideas come from understanding how trading actually happens, not what should happen in an efficient market. Index rebalance arbitrage exists because index funds are forced buyers on a known date; queue-jumping and latency arbitrage exist because of how exchanges match orders. These signals tend to be mechanical, testable against exact rules, and vulnerable the moment the rule changes (a rebalance methodology update, a tick-size regime shift).

The published literature, read skeptically. Academic finance publishes thousands of "anomalies" a year. Almost none survive out-of-sample, but reading the literature is still one of the highest-yield activities in research — not to copy a factor directly (it's likely arbitraged away by the time it's published) but to understand why a documented effect existed, then ask whether a related, less-picked-over version of the same mechanism might still be alive in a different market or at a different frequency.

Direct observation of frictions. The best ideas sometimes come from simply noticing something operationally clumsy: a fund forced to sell for reasons unrelated to value (a spin-off nobody wants to hold, a stock dropping out of an index), a corporate action that's mispriced because it's rare and few desks have built the machinery to handle it (see Handling A Corporate Action On A Live Book), a data field that's available but nobody's built a pipeline for yet.

Match the scrutiny to the source

The single biggest research mistake is applying the same casual level of scrutiny to every idea regardless of its source. An idea with a clean economic mechanism and a plausible reason to persist deserves testing across multiple markets and time periods to confirm it isn't specific to one dataset. An idea found by mining a large dataset for correlations, with no prior mechanism, deserves far more suspicion — with enough variables tested, some will correlate with returns by chance alone, and that idea needs a genuinely held-out sample and a strict multiple-testing correction before anyone believes the number (see p-Hacking and the Garden of Forking Paths). Researchers who source ideas mechanism-first and only then look for supporting data catch fewer chance patterns than researchers who mine data first and invent a story to fit whatever they found afterward — this ordering is the core of Hypothesis-First vs Data-First Research.

In practice

Good research groups keep a running list of candidate mechanisms and structural quirks worth investigating, refreshed by reading new academic work, watching for regulatory or market-structure changes (a new order type, a tick-size pilot), and post-mortems of trades that surprised the desk. New researchers are usually more productive spending their first weeks reading how existing signals were discovered than immediately mining new data — it teaches the pattern of "what a real mechanism looks like" that later lets them recognize one when they see it.

In interviews

Give a concrete example from each of the four sources rather than a generic list — interviewers are checking whether you can actually distinguish a mechanism-driven idea from a mined correlation, not whether you memorized the taxonomy. If asked how you'd generate a new idea today, describe starting from a plausible story about investor or institutional behavior, then testing whether the data supports it — not starting from a spreadsheet of correlations and working backward to a story.