The classifier’s estimated chance that a stock’s return falls below a fixed threshold — by default −5% within the horizon (default 63 trading days) — as a percentage. It is the mirror of buy probability: 50% is a coin flip; above 50% leans toward the drop, below 50% against.
The most common misread is treating sell probability as an expected loss. It isn’t. A 70% reading does not mean the model expects a 70% drop and says nothing about how far the stock might fall — only how likely it is to cross the −5% line at all. For the size of the move, read the expected-growth figure; sell probability answers “how likely,” not “how much.”
It is tempting to assume sell probability is just 100% minus buy probability, but they are two separate threshold questions — “will it clear +5%?” and “will it break −5%?” — with a wide middle band (between −5% and +5%) that belongs to neither. When a high sell probability sits next to a high buy probability, that is usually the model flagging a conflicted, high-uncertainty setup — the features disagree and the range of outcomes is wide — not that a big move is certain in either direction.
Example (illustrative): a ticker shows a sell probability of 65% at the three-month horizon. That reads as roughly a two-in-three chance of finishing more than 5% lower within about three months — and a one-in-three chance it doesn’t. It says nothing about whether the losing case is −6% or −30%. The figure is illustrative; the live ticker pages show each stock’s current reading when it is high enough to matter to the verdict.
The probability is calibrated — fitted on historical walk-forward folds so that, across many predictions, a stated 70% aims to line up with about a 70% observed rate of that drop. That is the goal, not a guarantee: calibration is measured in aggregate and per-ticker verification rests on small samples, so any single 70% is a best-calibrated estimate, not a promise this exact stock falls seven times in ten. Pair it with model confidence and the platform’s backtest accuracy.
No. It means roughly a 70% chance of falling below the return threshold (default −5%) within the horizon — the likelihood of that event, not the size of the fall. Read expected growth for magnitude.
No. They are two separate threshold questions — “will it clear +5%?” and “will it break −5%?” — with a wide middle band that belongs to neither, so they don’t sum to 100%. A high sell probability next to a high buy probability usually signals a conflicted, high-uncertainty setup, not a confident bearish call.
No. It is calibrated so a stated 70% aims to match about a 70% observed frequency in aggregate, but per-ticker verification rests on small samples — treat it as a best estimate, not a promise.
Browse all S&P 500 tickers to see this metric applied to individual companies.
Educational research only — not investment advice.