Conviction Score
A conviction score (sometimes "confidence score") is a numeric 0-100 estimate of how strongly an analyst — human or model — believes in a trade thesis. Top Tier Newswire's AI Top Trades model emits one per ticker per scoring run.
Scale
- 0-30 weak / speculative / single-signal — interesting but not actionable on its own
- 31-60 moderate — multiple signals pointing the same way, not overwhelmingly
- 61-85 strong multi-signal alignment — news, insider, congress, momentum, attention all aligned
- 86-100 overwhelming corroboration — reserved for unusually clean setups (rare)
Why it's bounded 0-100
The model's prompt explicitly anchors to a 0-100 rubric and is instructed not to cluster on round numbers (so you'll see 72, 78, 87 rather than always 70, 75, 80). The scale matters less than the relative ordering: the conviction-87 long is more important to the model than the conviction-72 long, regardless of whether you read 87 as "very confident" or "fairly confident".
Raw vs adjusted conviction
We track two values:
- conviction_raw: what the model emitted (its uncorrected output)
- conviction_adj: post-calibration value used to rank the board
The adjustment applies (1) per-conviction-bucket historical hit-rate calibration and (2) per-driver-set historical strength. If 70-bucket longs have hit 7-day 55% historically, future 70-bucket longs get a positive multiplier; if drivers in the call have weak realised win rates, the call gets dampened. Both adjustments are bounded so a single noisy bucket can't swing scores wildly.
The live board sorts on conviction_adj. The expanded row shows both numbers so you can see when the system has materially adjusted what the model said.