PMT DBStats Glossary
How we score and rank every take on Pardon My Take.
Card Stats
The four numbers on every speaker card
Fraction of judged takes that were correct. A .500 average means half the speaker's takes were right. Partially correct takes count as half.
AVG = (correct + 0.5 × partially correct) / total judgedLike batting average, but weighted by spiciness. A boring correct take barely moves the needle; a habanero-level correct take is a home run. Can exceed 1.000 for speakers who nail high-spice takes.
SLG = Σ(spiciness × outcome) / total judgedwhere outcome = 1.0 (correct), 0.5 (partially correct), or 0.0 (incorrect)
The speaker's average take spiciness, mapped to the Scoville scale for fun. Anchored at real pepper values: a level 2 maps to 2,500 SHU (jalapeño), a level 4 maps to 350,000 SHU (habanero).
SHU = 211.3 × (√140)^(level − 1)How much value a speaker adds above a “replacement level” speaker. Accounts for spiciness, conviction, outcome vs. expected base rate, and sarcasm. Positive means they're beating expectations; negative means they're falling short.
V = spiciness × conviction × (outcome − base_rate) × discountBase rate is calculated per spiciness level + take type (e.g. “how often are 4-spice predictions correct?”).
Sarcasm discount is asymmetric: sarcastic takes that beat expectations get full credit. Sarcastic takes that miss get discounted by (6 − sarcasm) / 5 — so a clearly sarcastic wrong take (5/5 sarcasm) only counts 20% against you.
Per-speaker TVOA is the average of all their take values, then shrunk toward zero via empirical Bayes (k=10).
Extended Stats
Additional stats shown on speaker profile pages
Borrowed from finance: measures risk-adjusted returns. Each judged take is treated as a bet where the wager is conviction × spiciness. Correct takes win the full wager, partially correct takes win 25%, and incorrect takes lose the full wager. The Sharpe ratio is the average payout divided by its standard deviation.
SHRP = mean(payouts) / stdev(payouts)Positive = consistently profitable on bold takes. Negative = the risk isn't paying off. A speaker with high AVG but low Sharpe is winning on safe takes and losing on spicy ones.
Career “winnings” treating every take like a bet. The wager on each take is conviction × spiciness — conviction is how much you'd bet, spiciness is the payout multiplier. A speaker who goes all-in (conviction 5) on a scorching take (spiciness 5) and nails it wins $25. If they whiff, they lose $25. Partially correct takes earn 25% of the wager.
BNKRL = Σ(conviction × spiciness × outcome)where outcome = +1 (correct), +0.25 (partial), or −1 (incorrect). This is a cumulative stat, so high-volume speakers will have more extreme values. Big positive = career take artist. Big negative = should probably stop making bold predictions.
The longest consecutive run of either correct or incorrect takes in the speaker's career — whichever is longer. Displayed as e.g. “14W” for a 14-take win streak or “22L” for a 22-take loss streak. Partially correct takes break both streak types.
A measure of category diversity from information theory. Low entropy means the speaker sticks to one or two topics (specialist). High entropy means their takes are spread across many categories (generalist).
ENTR = −Σ(p × log₂(p))where p is the fraction of takes in each category. A football-only speaker scores 0.0; a speaker evenly split across 16 categories scores 4.0.
Take Scores
The 1–5 ratings assigned to every individual take
How bold, contrarian, or hot the take is. A level 1 is room temperature (“the Chiefs are good”); a level 5 is scorching (“I think the 0-16 Browns make the playoffs”).
How committed the speaker is to the take. A hedged “maybe” scores low; a “I guarantee it, lock it in” scores a 5. Higher conviction amplifies both the reward for being right and the penalty for being wrong in TVOA.
How sarcastic or satirical the take is. PMT hosts love irony, so a take like “I think Kelvin Benjamin is in the best shape of his life” might score a 5/5. Sarcastic wrong takes are discounted in TVOA because the speaker wasn't serious. Sarcastic takes that turn out right still get full credit.
Correctness
How we judge whether a take landed
The take was right. Predictions that came true, opinions validated by events.
Partially right — the spirit was there but the details were off, or the outcome was mixed.
The take was wrong. Predictions that didn't happen, opinions contradicted by reality.
The take hasn't been resolved yet — future prediction still in play.
Pure opinions, jokes, or takes that can't meaningfully be judged right or wrong.
Empirical Bayes Shrinkage
Why small-sample stats are pulled toward the average
A guest who appears once and happens to get that take right has a 1.000 batting average — but that doesn't mean they're the best take artist ever. With only one data point, we can't distinguish skill from luck.
Empirical Bayes shrinkage solves this by pulling small-sample stats toward the population average. The fewer takes a speaker has, the more their stats are pulled toward the mean. High-volume speakers keep most of their raw numbers.
shrunk = (n × raw + k × global_avg) / (n + k)n = number of judged takes, k = shrinkage weight (higher = more conservative). For AVG and SLG, k is the median number of judged takes across all speakers. For TVOA, k is fixed at 10 and shrinks toward 0 (replacement level) instead of the global average.
Tier System
What the gold and red colors on speaker cards mean
After shrinkage, each stat is compared against the population of all speakers. Stats that are unusually high or low get colored:
Values between ±1σ are shown in neutral cream — within normal range. The exact thresholds update dynamically as more takes are processed.