Signal Sign Conventions
The simple but easy-to-botch rule for which direction a raw signal should point before it enters a backtest — a positive value meaning "buy" or "expect higher returns" consistently across every signal in a research pipeline.
A sign convention is the agreed rule that a positive signal value always means "expect this stock to outperform" and a negative value always means "expect it to underperform" — applied consistently across every signal a research pipeline touches. It sounds trivial, but signals arrive from wildly different sources (accruals, short interest, analyst revisions, momentum) and their raw, natural units point in inconsistent directions: high accruals historically predict lower future returns, while high momentum predicts higher future returns, so a "high value" is bullish for one and bearish for the other.
The fix is to flip signs at the point each signal is defined, so that downstream code — combining signals, computing information coefficients, sizing positions — can always assume "bigger number, more bullish" without checking each signal's origin story. Skipping this step is a common silent bug: a signal combiner or optimizer that adds a mis-signed component in with the wrong sign doesn't crash, it just quietly cancels out real alpha or actively works against it.
Worked example. A pipeline combines a momentum score (raw: positive = bullish, correct convention already) with an accruals score (raw: high accruals = bearish, so it must be negated before combining). If the accruals score is fed in unflipped, the combined signal effectively double-counts against high-momentum, low-accruals names — the two components partially offset instead of reinforcing — and a backtest of the combined signal shows a weaker information coefficient than either signal alone, a classic symptom traced back to one un-flipped sign.
Establish and document a single sign convention (positive = bullish) at the moment each signal is constructed; a signal fed into a combiner or optimizer with the wrong sign fails silently, quietly destroying or reversing alpha rather than throwing an error.
Related concepts
Practice in interviews
Further reading
- Grinold & Kahn, Active Portfolio Management, ch. 9