Researching Intraday vs Multi-Day Alpha

A researcher asked to look for "an intraday edge" and one asked to look for "a two-week edge" are not doing the same job with a different clock. The data, the backtest, and the bar for success are different at every step, and treating one like a faster version of the other is how projects quietly fail.

Different data, different backtest

A multi-day signal is built on daily closes, held through overnight risk, and evaluated against a universe that turns over slowly. The backtest can be a clean cross-sectional exercise: rank names today, hold for a week, measure next week's return. Costs matter but a rough model, a fixed number of basis points per side, is usually good enough to decide if the idea is alive.

An intraday signal lives inside a single trading session and dies or matures within minutes to hours. It needs tick or quote-level data, a realistic queue and fill model, and a cost estimate that accounts for the fact that you are often trading against the very short-term price move you're trying to capture, market impact is not a footnote, it is most of the return. A backtest that ignores the order book will pass signals that cannot survive contact with a real exchange.

What "it works" means changes

For a multi-day signal, a monthly information coefficient of 0.03 with modest turnover is a real result. For an intraday signal, IC per observation is nearly meaningless on its own, what matters is the IC accumulated across the thousands of independent bets a day produces, and whether the edge survives being executed in size without moving the price you're trying to trade at. A signal with a tiny per-trade edge can still be excellent if it fires often enough and costs stay below the edge; a signal with a large per-trade edge is worthless if it can only be captured a few times a day.

A worked comparison

Suppose a researcher has two candidate ideas. Idea A: order-flow imbalance predicts the next 5 minutes of return, with an IC of 0.06 measured tick-by-tick, but each signal is stale within 90 seconds. Idea B: analyst-estimate revisions predict the next 10 trading days, with an IC of 0.02 measured daily. Idea A needs a fill-probability model and realistic latency assumptions before the 0.06 means anything, spread alone can erase it. Idea B can be sanity-checked in an afternoon with daily bars and a flat cost assumption, because overnight and multi-day holding smooths out the microstructure noise that would dominate Idea A's evaluation.