Win probability and upsets: what your leaderboard isn't telling you

A leaderboard answers one question: who's on top? Useful, but static. The more interesting question is asked before every match: what should happen here?

That's win probability, and it's hiding in the ratings you already have.

The gap is a forecast

Rating systems like Elo aren't just rankings - they're prediction machines. The gap between two ratings converts directly into odds:

Rating gap	Favorite's win chance
0	50%
50	57%
100	64%
200	76%
300	85%
400	91%

So when the office #1 (1450) plays the #5 (1280), the math says 72/28. Not a coin flip, not a foregone conclusion - a forecast. And like all good forecasts, its value is in being checkable: over a season, players with a 70% pre-match probability should win… about 70% of those matches. If they don't, the ratings adjust until the forecasts come true. That feedback loop is the entire magic trick.

(Glicko-2 and TrueSkill make the same forecast with extra honesty - they track how certain the system is about each rating, so a newcomer's 60% is appropriately humbler than a veteran's.)

What counts as an upset?

There's no official threshold, but a useful one: the winner had less than a 30% pre-match chance. Below 30%, you're past "competitive match, either could win" and into "the table expected something else."

Why does this deserve a label at all? Because upsets are information-rich. In Elo terms, a sub-30% win moves more rating points than any routine result - the system treats it as the strongest possible evidence that the loser was overrated, the winner underrated, or both. One genuine giant-killing is worth roughly three expected wins.

And culturally, upsets are the entire product. Nobody retells the story of the favorite winning 11–6. The new hire taking down the office champion on a Tuesday afternoon - that story survives the week, fuels the rematch, and drags three spectators to the next match. A leaderboard that flags those moments ("⚡ Upset - pre-match win chance: 22%") is doing the office's storytelling for it, which is exactly why we built that into Trofeo's match feed.

Reading your own probability

Seeing "28%" next to your name before a match changes how the match feels, in a good way:

• As the underdog, you're freerolling. Lose, and you've lost almost nothing - the forecast already priced it in (a 28% underdog who loses gives up only a few points). Win, and it's a heist: maximum points, a flagged upset, a story.
• As the favorite, the number is a tax notice. A 72% favorite gains little by winning and bleeds by losing. That pressure is real, and it's fair - it's the cost of being on top, and it's why staying #1 is harder than getting there.
• As a matchmaker, probabilities tell you which games to schedule. The best office rivalries live in the 40–60% band: genuinely uncertain, maximally fun. If your lunchtime fixture has drifted to 85/15, it's time to find a new nemesis - or play doubles with handicapped pairings.

The honest caveats

Win probability is a model, and models have edges worth knowing:

• Small samples lie. A rating built on six matches is a sketch, not a portrait. Uncertainty-aware systems (Glicko-2, TrueSkill) handle this by widening the error bars; either way, treat early-season probabilities loosely.
• Form is invisible. The model knows your rating, not that you slept four hours or that the office is using new balls. Day-to-day variance is exactly why 76% favorites lose a quarter of the time - and why the matches still need playing.
• Probabilities are not promises. A 91% favorite losing isn't the model being wrong; it's the 9% happening. Over a season, the calibration is what matters.

The takeaway

Ratings rank the past. Probabilities frame the future. Together they turn a list of names into a season-long narrative - favorites defending, underdogs hunting, and every match carrying a number that someone gets to prove wrong.

Check the forecast before your next office match. Then go ruin it.