The conviction score’s downside check, and the reason a promising-looking stock can still be refused. It ignores what the model expects to happen and looks only at what happens if things go badly: how far the stock falls in the model’s bad-case scenarios. A shallow worst case scores well; a deep one scores zero and caps the entire signal.
The model does not produce a single expected price — it produces a whole range of possible outcomes, and from that range it publishes a 90% confidence band: a lower edge and an upper edge that the price is expected to finish between in all but the more extreme cases. The downside bound this component scores is the LOWER EDGE of that band, expressed as a percentage move from today’s price.
So a downside bound of -8% means: in the model’s world, the bad case for this stock over the horizon is roughly an 8% fall.
One honest detail worth knowing: that lower edge is not simply a headcount of the worst simulated paths. The band is calibrated — where backtesting showed the model’s raw bands were too confident, they are widened, and never narrowed below the model’s own uncertainty. How well the bands have actually held up is published on our calibration page rather than asserted here.
Three steps, deliberately coarse. A downside bound shallower than -5% scores 1 — full marks, the bad case is mild. Between -5% and -15% scores ½ — a real but survivable drop. Worse than -15% scores 0 — the model’s own bad case is a serious loss.
Example (illustrative): two stocks both have the model pointing firmly upward. Stock A’s downside bound is -4%, so tail risk scores 1. Stock B’s is -22%, so tail risk scores 0 — and because this is a gate component, Stock B’s entire conviction score is capped at 3.0 out of 10 regardless of how bullish everything else on its page looks. Same upside thesis, opposite verdicts, and the difference is purely the shape of the downside.
Because losses and gains are not symmetric in their effect on your account. Halve your money and you need it to double just to get back to where you started. A strategy that takes attractive-looking bets with deep tails will look brilliant for a while and then hand back years of returns in a month.
So tail risk is one of the two components that can veto everything else. If it scores 0, the conviction score is capped at 3.0 out of 10 — no technical indicator, no trend, no momentum reading can rescue it. The other veto component is quantum direction. The platform would rather miss a good trade than take a bad one: a missed trade is cheap, a bad trade is expensive.
Use it as the sanity check on your position size. A stock scoring 1 has a mild modelled bad case; a stock scoring ½ has a real one, and you should size the position so that the bad case is an inconvenience rather than a catastrophe. A stock scoring 0 is one the platform has already refused to publish a BUY on — treat that as the answer, not as a challenge.
No, and this matters. It is the lower edge of the model’s 90% confidence band, not a floor — the model itself expects prices to finish below it some of the time, and the band is only approximately calibrated. It is a bound on the ordinary bad case, and it offers no protection against a crash, a fraud or an overnight gap.
Because the two components ask different questions. Direction asks which way the model leans; tail risk asks how bad it gets if the model is wrong. A forecast can lean upward and still carry a deep downside tail — that combination is exactly what this component exists to catch and refuse.
Max drawdown is historical — the worst peak-to-trough fall the stock actually suffered in the past. Quantum tail risk is forward-looking — the bad-case outcome in the model’s simulated paths over the forecast horizon. They often disagree, and it is worth looking at both.
Browse all S&P 500 tickers to see this metric applied to individual companies.
Educational research only — not investment advice.