Read probabilities and choose thresholds

A probability is not a signal until you decide where to act. Here is how to pick that line and check it against what really happened.

Start from the base rate
Every outcome has a normal frequency. A 5% drawdown within a month happens often; a volume day above twice the 20-day average is rare. So 0.18 for a volume spike can be three times the usual odds, while 0.50 for a drawdown may be ordinary. Compare an answer with what the model says for a quiet, typical stock.
Pick a threshold per question
Act only where the probability is clearly away from the base rate, for example drop5 >= 0.6 to cut size, or up_week >= 0.6 to add to a watchlist. Higher thresholds fire less often and are right more often.
Check against real outcomes
Open Examples → Gap fill or Paraphrase check: they come with what actually happened. Run them and use the Outcome buttons to see which answers were closer.
Use confidence for choice and score
choice: (p_max − 1/n) / (1 − 1/n). score: 1 − (expected distance from the most likely level ÷ that distance under a uniform guess), floored at 0. Route low-confidence answers to a person instead of acting on them. yes_no has no separate confidence: the probability is the belief.
Turn it into a rule
Put the threshold in an action's when so it runs on our side: see Send a Slack alert.

What's next

Create an account Open the Playground Everything here also lives in the console, next to your keys and usage.