Calibrated Probabilities vs Buy/Sell Ratings: What the Difference Costs You
2026-11-12 ยท hedgewing.ai Research
Most stock research ends in a label: Buy, Hold, Sell, Strong Buy, a 1-to-5 rank, a score out of 10. Labels are easy to read and easy to compare, and that is exactly their problem. A Buy on a stock with a coin-flip outlook and a Buy on a stock with a genuine edge look identical on the page, and nothing about the label tells you which one you are holding.
What a rating actually encodes
A rating is an ordinal signal. It says this stock ranks above that one, or that the analyst leans positive. It does not say how often stocks with this rating went up, by how much, or over what horizon. That makes it impossible to score: if a Buy-rated stock falls, was the rating wrong, or was it the one-in-three case the rating always implied? Without a stated probability there is no way to tell, and a signal that cannot be wrong in a measurable way cannot be improved either.
What a calibrated probability encodes
A calibrated probability makes a falsifiable claim: of all the times the model says 62% for a five-day rise, about 62% should resolve higher. That is checkable with a reliability diagram (predicted probability on one axis, observed frequency on the other) and scoreable with a Brier score. Calibration is separate from accuracy. A model can be right 58% of the time and still be badly calibrated if it reports 90% on every call; a model that reports 55% when it means 55% is more useful even if its raw hit rate is the same. The glossary entries at /glossary/calibrated-confidence and /glossary/brier-score cover the mechanics.
Why the difference matters for position size
Position sizing is where the two signals diverge in practice. With a rating, every Buy gets the same size because there is nothing to scale by. With a probability, the size can follow the edge: a 68% call with unanimous model agreement can justify a position near your per-trade risk cap, a 54% call with a split vote justifies a small one or none. Expected value makes this concrete. A 60% chance of a 4% gain against a 40% chance of a 4% loss is a 0.8% expected return before costs; a 52% chance on the same payoff is 0.16%, which a 0.2% round-trip cost turns negative. The rating would have said Buy in both cases.
How to check whether a probability is real
Three questions separate calibrated probabilities from decorated ones. First, were the probabilities validated against outcomes the model did not train on, and can the provider show the reliability curve? Second, do hit rates rise with stated confidence, so that the 70% bucket really resolves more often than the 55% bucket? Third, does confidence fall when it should, around earnings dates, in volatility spikes, and when the underlying models disagree? A number that sits at 85% regardless of conditions is a rating wearing a percent sign.
How hedgewing.ai reports it
hedgewing.ai attaches a calibrated confidence to every forecast horizon and, on Pro, shows the per-model breakdown behind it so you can see whether the four architectures agree. The platform makes no buy or sell recommendations; it reports a probability and the risk analytics to act on it. Backtests shown on the platform are on-demand, in-sample, and illustrative rather than a live track record, and no probability, however well calibrated, guarantees an outcome. This article is educational and is not investment advice.