Sentiment Signals

A sentiment signal turns mood into a number. How positive is the tone of the news about a stock? Are earnings-call transcripts hedging and defensive or confident? Is social-media chatter turning euphoric? Are analysts revising estimates up or down? Is the options market pricing fear? Each of these is a read on what investors feel, and feelings move prices before fundamentals catch up — sometimes helpfully, sometimes as a warning that a move has gone too far.

The reason sentiment carries information at all is that markets aren't perfectly rational in the short run. Attention is limited, reactions are slow and then overdone, and narratives spread. Tetlock's classic result — that pessimistic media tone predicts short-term downward pressure and a later reversal — captures the two-speed nature of the whole field: sentiment drifts in the middle and reverses at the extremes.

Building the signal

Most modern sentiment signals come from text. A model scores a document — a news story, a tweet, a 10-K, a call transcript — on a scale from negative to positive, often using finance-specific dictionaries (generic "liability" or "tax" are negative in everyday language but neutral in filings, the point Loughran and McDonald made). Aggregate the scores for each name and you get a cross-sectional signal:

$$S_i = \frac{(\text{positive mentions}) - (\text{negative mentions})}{\text{total mentions}},$$

where $S_i$ is the net sentiment for stock $i$, running from $-1$ (uniformly negative) to $+1$ (uniformly positive). You then rank $S_i$ across all names and, in the trending regime, go long the most positive and short the most negative — the same Signal Construction pipeline any factor uses.

Worked example: news tone as a cross-sectional tilt

You score every S&P 500 stock's news over the past week. Company A has 40 positive and 10 negative mentions; Company B has 12 positive and 28 negative:

$$S_A = \frac{40 - 10}{50} = +0.60, \qquad S_B = \frac{12 - 28}{40} = -0.40.$$

Rank all 500 names by $S_i$; A lands near the top decile, B near the bottom. You go long the top decile, short the bottom, dollar- and sector-neutral. Backtested, this kind of news-tone signal has earned a modest but positive spread over the following one-to-two weeks before decaying — consistent with under-reaction that the market slowly corrects. Crucially, if A's score were an off-the-charts +0.95 driven by a frenzy, you'd treat it with suspicion rather than pile in: that's the extreme zone where the drift flips to reversal.

Where sentiment misleads

Mood is the noisiest input in the quant toolbox, and it fights back.

• It's reflexive and gameable. Once a signal is known to move prices, actors flood the channel — pump-and-dump posts, bot armies, press-release spam. Social-media sentiment is especially easy to manufacture.
• Volume, not just tone, matters. A mildly positive story that everyone is reading can move more than a wildly positive story nobody sees. Weight by attention, or the score misleads.
• Extremes flip without warning. The contrarian zones are real but their boundaries move; sizing a fade to "this is extreme" has blown up many a short into a squeeze.
• Fast decay. Sentiment edges live for days, not months — high turnover, so Transaction Costs eat naive implementations quickly.

Sentiment sits right beside Alternative Data Signals in the modern signal stack: one reads what people say and feel, the other reads what they do. Both are timely and orthogonal to price, both decay fast, and both demand the same skepticism — because the easiest thing in the world is to find a mood that fit the last rally and won't fit the next one.