Demeaning: Universe, Sector or Industry?

"Demean the signal" sounds like a one-line preprocessing step, but the question of which mean is where a lot of the actual research judgment lives. Subtract the average across the whole universe, and you keep sector bets in the signal. Subtract the sector average, and you kill them. Subtract a narrow industry average, and you may kill real information along with the noise.

What each level of demeaning does

Universe-wide demeaning compares every stock to the average of the entire investable set. If tech has been strong all year, a tech stock with an ordinary momentum score will still look attractive relative to the whole market, because the whole market includes lagging sectors. The signal, left this way, is really making two bets at once: a stock-picking bet and a sector-rotation bet.

Sector demeaning compares each stock only to others in its own sector. The tech stock above is now judged against other tech stocks; if tech as a group is strong, that strength is subtracted out, and what's left is purely "is this stock unusually strong for a tech name." The sector bet disappears; only the stock-selection bet remains.

Industry demeaning narrows the comparison group further — semiconductors versus semiconductors, not semiconductors versus all of tech. This removes even finer common movements (a whole sub-industry re-rating together), but the narrower the group, the fewer stocks are in it, and the mean itself becomes noisier and more sensitive to one or two names.

Level	What's removed	What's kept	Risk
Universe	Nothing but the overall market drift	Sector bets + stock bets	Signal doubles as a sector-timing bet, often unintentionally
Sector	Sector-wide moves	Pure stock-selection within sector	Cleanest for a pure stock-picker
Industry	Sub-sector moves too	Very narrow within-industry differences	Small groups, noisy means, real information can get demeaned away

A worked judgement call

A researcher builds an earnings-revision signal and finds it makes money but almost entirely through being overweight semiconductors during a multi-year sector re-rating. Universe demeaning alone did not catch this, because the whole tech sector's revisions were rising together and every tech name looked attractive relative to laggard sectors. Moving to sector demeaning removes that common component; the signal's return, once retested, drops sharply, revealing that most of the original "alpha" was really a bet that semiconductors would keep re-rating — a bet the researcher had not intended to make and could not defend as a repeatable edge.