PowerScore boosts and penalties

Boosts and penalties sit on top of the five signals rather than replacing any of them. Two are computed by the package itself and returned separately from computePowerScore. A third is folded into total automatically.

Scoring opportunity boost

computeScoringOpportunityBoost(game) returns extra points for a scoring threat the score itself doesn’t show yet. It returns 0 unless game.status is 'in' and isPlayFrozen(game) is false. A frozen game holds its runners, or its down and distance, in place with nothing able to happen.

Baseball and softball, by runner count, from scoringOpportunityBaseRunnerBoosts:

Football, gated on how close the game already is, from scoringOpportunityRedZoneBoost and scoringOpportunityRedZoneFringeBoost:

The base value is then weighted by down, from redZoneDownMultipliers:

A base value of 0 stays 0 no matter the down. An unconditional boost on top of a blowout would undo what closeness and late-game pressure already scored correctly low. Goal-to-go only raises the multiplier on 4th down, since on an earlier down it describes the odds of a score rather than what decides possession.

This boost is not part of total. Score a game from live data shows how to add it to one.

Win probability balance

computeWinProbVarianceScore(winProbHistory) returns a modifier from -scoreWinProbVarianceMax to +scoreWinProbVarianceMax (±5). Unlike the scoring opportunity boost, computePowerScore folds it into total on its own, when a winProbabilityHistory argument is supplied.

It needs at least minDataPoints (5) finite values in the array. Anything else returns undefined, and the result’s winProbabilityVariance field is left off entirely rather than set to 0. Non-finite entries (NaN, Infinity) are dropped before that count is checked.

The modifier is the mean absolute distance of the supplied values from 0.5, mapped onto the ±5 range:

Despite the name, the modifier measures average distance from 50%, not variance. A win-probability line that oscillates between 10% and 90% scores the same penalty as one held steadily at 90%. Both average the same distance from the midpoint.

The win-probability modifier applies after the stall penalty, not before it. computePowerScore deducts the stall penalty from the five signals first, then adds the modifier on top of that reduced subtotal.

Stall penalty

A flat deduction applies to the five signals’ subtotal when a game’s clock has stopped moving for several consecutive polls. The deduction keeps ArenaSwap from switching to a game sitting in a commercial break or a timeout.

computePowerScore doesn’t detect a stall itself. It takes a stallCount argument, and the caller counts how many consecutive polls the clock hasn’t moved. Score a game from live data shows how.

stallPenaltySteps, checked highest threshold first:

The deduction comes straight off the signals subtotal rather than scaling it. A game whose five signals sum to 79 lands at 54 under the heavy penalty, not 55. signalsSubtotal on the result holds the pre-deduction sum, so a breakdown can show what the penalty removed. The win-probability modifier is unaffected, since it’s added after the deduction, not before.

Fields the package doesn’t compute

PowerScoreResult also carries favoriteBonus, favoriteTeamCount, gameBoost, and postseasonBoost. normalizePowerScoreResult clamps and passes these through when supplied. computePowerScore never sets or adds any of them. They exist so a consumer can attach its own boosts and still get back one normalized result. Their exact shape and optionality are in PowerScore’s TypeScript types. How ArenaSwap itself turns a viewer’s saved preferences into these numbers is covered in its own docs, not this package.