⚡ How the Upset Meter works

Fitted to our games, not borrowed

We took every completed, rated, handicap-free match in this club's history and asked one question: when two teams with a known rating gap play, how often does the favorite actually win?The answer became the curve the meter runs on. It is not DUPR's algorithm and it never talks to DUPR — it's fitted to what happens on our courts.

1. Win probability

For a team rating gap g(average of the two partners' ratings, favorite minus underdog):

P(favorite wins) = 1 / (1 + 10^{−g / 0.77})

The 0.77 isn't a guess — it's the value that best matched our real results. Check the model against the club's actual history:

If the meter were noise, those two columns wouldn't line up bucket for bucket.

2. Predicted score

The same history shows the losing score falls about 2.8 points for every 1.0 of rating gap, starting from ~6.6 in an even game:

loser's score ≈ 6.6 − 2.8 × gap

(scaled for 15- and 21-point games). Worked example: a 0.25 gap predicts roughly 11–6 with the favorite winning 68% of the time. That's also why most forecasts read 11–5 or 11–6: real pickleball scores compress toward the middle — the win percentage is the number that separates matchups.

3. Ratings that learn

A rating card can lag reality — someone improves fast, or comes back rusty. So the meter tracks each player's results against what their rating predicted, and solves for the adjustment d that makes the math match the record:

Σ P(gap_i + d) = actual wins

Worked example: a 3.9-rated player wins 8 of their last 10 even matchups. A true 3.9 should win about 5, so the equation solves to d ≈ +0.46 — but small samples get shrunk by n / (n + 10), so at 10 games only half counts. The meter quietly treats them as a ~4.13until the results say otherwise. It works both directions (cold streaks adjust down), it's capped at ±1.0, and it fades back out when results return to normal. A mis-rated player stops generating fake "upsets."

4. New partners get the benefit of the doubt

A first game with a brand-new partner is a gamble — styles, communication, who takes the middle. Statistically, that game is confounded with unknown chemistry, so it can't be pinned on either player's skill. Some rating systems ignore this and let one ugly loss with a stranger crater a long record. We weight it instead:

game weight = m / (m + 2),  m = games this exact duo has played

Your first game with a new partner counts one-third toward your adjustment, the second counts half, and the discount fades as the partnership gets identified. And once a duo has 3+ rated games together, their measured chemistry— how far they run above or below the sum of their ratings, capped at ±0.5 — is folded into that pair's forecasts. Established teams get priced as teams, strangers get grace.

5. Championship odds

Season pages take this one step further: once a week is in the books, we simulate the rest of the season 2,000 times — every game sampled from the curve above, including ladder court movement and court-weighted points — and count how often each player finishes first. The odds move every week as real results replace simulated ones.

The honesty rules

• Predictions use the ratings teams carried when they played — no hindsight.
• Handicap games are excluded — the spot exists to even the odds.
• Rec play nights never move anyone's adjustment. Test a new paddle without re-pricing yourself.
• One bad game with a first-time partner can never tank your number — new-partner games are discounted until the pairing has a track record.
• A 75% favorite still loses one game in four. The meter doesn't call winners — it prices upsets. Beating it is the fun part, and it lands you on your league's Biggest Upsets board.
• Anyone who'd rather play without numbers in their head can switch predictions off on their profile. The club still tracks upsets — they just don't see the odds.