How to use a sentiment signal without overfitting it to last year’s crash

CrowdMood does not sell a guaranteed equity edge. What we sell, and what we publish for free, is a way to keep household mood in an operating system without turning it into a lucky chart.

One hypothesis per rule

A defensible use is a single sentence. Example: “When our U.S. current-conditions z-score stays below −1 for three prints, cut discretionary ad spend by 10% until it mean-reverts through −0.5.” That sentence can be wrong. It can also be tested. A dashboard with fourteen arrows cannot. If your signal vendor’s PDF has more rules than pages of out-of-sample results, you are looking at a content site with a price list.

Costs, lags, and the survey calendar

Surveys are dated. A Friday print describes interviews that may have ended days earlier. By the time you trade the open, the information is in other prices. That does not make the print useless for a retailer whose open-to-buy meeting is Tuesday. It does make a same-day S&P strategy a different, harder object. We document lags in the methodology page because hiding them is how backtests get pretty.

Include transaction costs, missed fills, and the fact that you will not size like the backtest when the rule is losing. Sentiment strategies that look brilliant in 2022’s inflation scare often fade when inflation is boring. Regime labels are not free. They are extra parameters.

What we ship

The public indices are scaled around 50, with 40–60 as the ordinary band. They are not buy/sell lights. Licensed CSV files add z-scores, slopes, revision flags, and a desk comment that may read “stand down.” API users get the same fields.

If you build your own overlay, keep a written log of every change to the rule, with the date you changed it. Future you will try to forget the tweaks. That log is the real edge: institutional memory, not a sharper RSI of Michigan.

Educational use is welcome. Scraping our public tape into an undisclosed paid product is not. The terms page says so in quieter language. Research is a commons only if people do not enclose it on the way out the door.