Confidence Meter

The simplest way to read a pick: a 0-100% bar. It's built from the same underlying score as the verdict badge below it, just rescaled into one number — the meter and the badge will always agree, they're two views of the same judgment, not two different opinions. Above ~75% and green is the model's best-supported picks; below ~30% and red means it doesn't have real conviction here, treat it as a SKIP.

Predicted Kills

The model's point estimate for this player's combined kills across Map 1 + Map 2 of their next series (PrizePicks' "Maps 1-2 Kills" convention) — the headline number the dashboard leads with.

80% Range

A confidence interval: the model expects the real outcome to land inside this range about 80% of the time. A range that fully clears a betting line (entirely above or below it) is a stronger signal than one that overlaps it, since the model isn't just leaning a direction — it's saying the whole plausible outcome band sits on one side.

Model Blend

Predicted Kills is a weighted blend of two independently-trained engines — an 11-model stacked regression ensemble ("V2") and a Poisson/Negative-Binomial two-stage model ("Ultimate"). Each gets a vote proportional to its own validated accuracy (inverse cross-validation error), and disagreement between the two widens the uncertainty band — two different architectures agreeing is a stronger signal than either one agreeing with itself.

Typical Miss (Position CV MAE)

On average, how many kills off the model's guess tends to be for this role specifically, measured on real held-out data — not a marketing number. Lower is better.

Track Record (Direction Accuracy)

How often the model's OVER/UNDER call was actually correct, validated on data the model didn't train on. The relevant benchmark is breakeven — roughly 52.4% for standard even-money bets — so anything meaningfully above that is a real edge, not just "looks impressive."

Model Agreement (Ensemble Agreement)

How tightly the individual sub-models (11 for V2, plus Ultimate) agree with each other on this specific player. Tight agreement (small ± spread) is a stronger, player-specific confidence signal than the position's generic accuracy number alone.

Classifier & Meta-Learner

Two extra opinions layered on top of the base prediction, only shown once you type in a line. The Classifier is a second model trained specifically to answer "will this go over THIS line?" — it never decides the direction (the main prediction always does that), it just gives a calibrated percentage. The Meta-Learner is a third model that's learned WHEN the first two tend to agree and actually be right — think of it as a referee checking whether this specific combination of signals has historically been trustworthy. Both show up as extra lines in "Why This Verdict" and feed into the Confidence Meter.

Context Chips

Signals the model already learned and had baked into its prediction, surfaced so you can see WHY it landed where it did — hot/cold streak, team pace, patch trends, rest days, side, league tier. Chips that have a real directional lean are direction-aware: their color reflects whether they support or contradict the CURRENT pick (OVER or UNDER), and their label always states what they favor in plain text (e.g. "favors OVER") so the color never needs explaining.

Verdict Badge

A synthesized read on a specific PrizePicks-style line, combining edge size (how far the prediction sits from the line), whether the 80% range clears the line, sub-model agreement, this position's accuracy vs. breakeven, the classifier/meta-learner (if a line is entered), and how many context chips support vs. contradict the call. Strong Pick = everything lines up. Lean = a real but modest edge. Not a Lock = genuinely uncertain. SKIP = the model doesn't have real conviction here — thin edge, wide disagreement, or below-breakeven accuracy for this position. A SKIP is not a bug; declining to call an uncertain pick is the point.

Live Now / Watch

On the Live Odds page, any match currently being played shows a pulsing LIVE badge. Click "Watch" on it to play the official broadcast right there on the page (click "Hide" to close it) — no need to leave the app or go find the stream yourself.

Tier 1

Live Odds and its refresh are scoped to the four major regional leagues (LCK, LPL, LEC, LCS/LTA) plus international majors (Worlds, MSI, First Stand) — not academy, challenger, or regional-qualifier leagues. This keeps the free odds API's hourly request quota sustainable and keeps the feed focused on the matches that actually have betting markets worth checking.

Edge Finder / Model Total

Since no sportsbook prices individual LoL player kill props yet, Edge Finder compares the model's own player-by-player predictions, summed to a team total, against the closest real market that exists: a book's team-total-kills line for the match. "Model Total" is that sum; the table below it is what the book(s) are actually offering.

Confident-Pick Accuracy (Model Report)

On the Model Report page, below each position's base accuracy you may see a second, green bar: the accuracy of ONLY the picks where edge was meaningful AND the meta-learner was confident (≥55%). This is real, honest walk-forward data (not the model grading its own homework) and it's consistently higher than the base number — that gap is the whole point of the classifier + meta-learner: they're built to find the subset of picks worth actually trusting.