Neutralising Sentiment Signals Against Size and Sector

A raw sentiment score built from news text is rarely a clean signal on its own — small-cap stocks tend to get sparser, noisier coverage than large caps, and certain sectors (say, biotech) generate more emotionally charged headlines than others regardless of whether anything meaningful happened. If a strategy trades raw sentiment directly, it ends up making an accidental bet on size or sector rather than the sentiment content itself, and that hidden bet can dominate the signal's apparent performance.

Neutralising a signal means regressing it against known characteristics — market capitalization, sector membership, recent volatility — and keeping only the residual, the part of the sentiment score that isn't explained by those characteristics. What's left is a cleaner bet on the news itself, isolated from exposures the strategy didn't intend to take.

Worked example

Suppose small-cap stocks average a sentiment score of +0.4 purely because their headlines run hotter, while large caps average 0.0. A stock in the small-cap bucket with a raw score of +0.5 isn't unusually positive at all — it's roughly in line with its peer group. Neutralising subtracts the group average, leaving a residual of +0.1, which is the genuinely informative part of the signal.