Chasing Down Unexplained P&L

Run a proper attribution and you'll break the day's P&L into pieces — sector, factor exposures, stock selection, carry, hedges. Add them up and they almost never equal exactly the actual P&L. There's a leftover, usually called the residual or the unexplained P&L. A small residual is normal; every attribution model is an approximation. A residual that's large, or that shows up in the same direction day after day, is a signal that your model is missing something real, and it's worth chasing down before it costs you.

Where residuals come from

Three sources account for almost all of them. First, timing: your factor exposures and your attribution model are usually measured at a snapshot (last night's close), while trading happens intraday, so a name that moves sharply between the snapshot and your actual trade time shows up as unexplained. Second, missing risk factors: if your model has ten factors and the book has meaningful exposure to an eleventh — say, a commodity-sensitivity factor nobody's modeling — every day that factor moves, you get a residual. Third, plain data errors: a stale price, a wrong corporate-action adjustment, a position booked with the wrong sign.

Worked example

The book's total P&L for the day is +$180,000. Attribution assigns +$95,000 to stock selection, +$40,000 to sector tilts, +$15,000 to the factor model's explained beta exposure. That sums to +$150,000, leaving a residual of +$30,000 — about 17 percent of the day's total.

You check the usual suspects. Timing: no unusual intraday moves in the largest positions. Data: prices all confirm against a second source. That leaves missing factors, so you regress the residual against a list of candidate exposures the model doesn't carry and find it lines up almost exactly with the day's move in high-yield credit spreads — the book has an unmodeled sensitivity to credit through two names that are effectively credit-sensitive but get bucketed as plain equity. That's not noise; it's a real, recurring exposure the risk model has been blind to.

When to stop chasing

Not every residual is worth this effort. If it's small relative to daily P&L (a rough rule of thumb many desks use is under 10 percent) and doesn't persist in the same direction across a week, it's very likely noise — model approximation error and timing mismatch, nothing to fix. The signal that something's wrong is persistence: a residual that's consistently positive or consistently negative for two or three weeks running is not random error, because random error doesn't have a sign.