Not a Bet. A Goal Differential With a Jump Layer.

Three things are being compared this month, and they are not variations on a theme. CME's new contracts are cash-settled futures on a cumulative, additive statistics index built from official NHL data, administered by FutureSports under an exclusive league data agreement signed August 4 and IOSCO-aligned governance [1][2][3]. Kalshi's sports products are fully collateralised binary contracts on discrete outcomes, which carried $2.72 billion of sports contract volume in the week to August 30, 88.6% of the sports category measured against Polymarket Global and 90.3% of all combined contract volume that week and which the Ninth Circuit held on August 28 are not swaps at all, splitting with the Third Circuit's April decision that they are [4][5][6]. Adjacent's NFL Team Indices are price-derived: the running mark of an equal-dollar basket across four 25%-weighted sleeves — Super Bowl, conference, division and team total wins — held in Kalshi and Polymarket outcome markets and chained multiplicatively from a base of 1,000 [7].

None of the three references another for settlement, data, or governance. FSPI settles to NHL statistics as processed by FutureSports; Kalshi settles to game results under its own rulebook; Adjacent settles to the prices of Kalshi and Polymarket contracts. The only relationship is economic. All three are functions of how good a team is and how its games go, observed at different points on the chain from stat production to realised outcome to market-implied probability of the outcome, and with different time orientation: backward-looking and cumulative, terminal and binary, forward-looking and continuous.

That absence of linkage is the design. CME's binary sports ambitions with FanDuel unwound over eight months. FanDuel Predicts launched on December 22, 2025 with CME listing financial contracts and sports across four leagues in states without legal online sportsbooks. In June 2026 FanDuel added Crypto.com and OG Prediction Markets alongside CME rather than in place of it; then on August 5, 2026 Flutter told the market that, “in coordination with CME,” all FanDuel Predicts sports and novelties contracts would move to Crypto.com. Duffy’s public objection — that sports prediction markets are largely gambling and that small parlays invite manipulation — was made plainly on the July 2026 earnings call. CME kept its 51% stake in the venture and lost essentially all of the volume, the retained non-sports contracts having been reported at under 1% of the app’s activity, three weeks before it announced the NHL index futures [8][9]. The FSPI product is CME entering sports through the index door instead — filed under its equities complex, drafted so that the index "does not settle, terminate, or resolve upon the occurrence or non-occurrence of any single game outcome" [2], and scheduled to list one month after Assad. The trade press will compare it to Kalshi anyway. The interesting question is what it actually is, and the methodology is public enough to answer that with data rather than adjectives.

The published guide specifies a 55-row points-attribution table [2], of which 53 are performance constituents and the last two are the seasonal adjustment factor constants. Every team starts each season at 7,500. In-game statistics add or subtract fixed multipliers — an even-strength goal +12 (allowed −12), a power-play goal +10, a short-handed goal +15, a shot on goal +1.5 (opponent's −1.5), a save +0.75, a takeaway +2.5, a giveaway −0.8, a blocked shot +0.8, a faceoff win +0.5, a hit +0.1, penalties −1 to −10. After each game a milestone layer is applied: win +10, regulation loss −10, overtime loss −5, shutout ±50, a game-clinching goal +30, six-plus goals +35, exactly five +20, exactly one −20 (each mirrored for goals allowed), save percentage at or above .940 +10 and at or below .850 −12.5. Month-ends award ±125 for most and fewest goals, blocks, takeaways and giveaways; the regular season's end awards ±250 for the same categories plus first and last in each conference; the Stanley Cup is +750 with the 1.5× postseason factor applied to everything, so +1,125. Ties pay every tied team in full. The official close prints at 9:00 a.m. Central on T+1, prior closes are not restated except under the separate index-maintenance and error-management document, and there is no floor — the index may go negative.

Kinetic Alpha rebuilt the index from the NHL's official play-by-play feed for six seasons, 15,410 team-games, applying every constituent as written. The result is in Figure 1.

The long constituent list reads like a possession index. In variance terms it is not. Because the shot and save terms are nearly symmetric between the two teams in a game — your shots on goal are the opponent's saves plus your goals — they cancel almost entirely, and the per-game increment collapses to roughly 0.75 × shot differential + 12.75 × goal differential ± 10 for the result, plus milestones. Decomposing the realised variance makes the point sharper.

So the honest description is a cumulative goal-differential index with a win/loss kicker and a jump layer. It is closer to outcomes than the constituent table suggests, but it is not an outcome contract: it is path-dependent, it accumulates, and it never resolves on a single game. The two teams in a game are almost perfectly anticorrelated (ρ = −0.96), and the sum of their two increments has a mean of +17.6 and a standard deviation of only 20, so a long/short pair in the same game is nearly riskless for that day; away from a shared game, cross-team correlation is essentially zero.

CME will presumably margin FSPI through SPAN 2 at 99%-plus coverage over a one- or two-day horizon, as it does for other cash-settled index futures. The contract is $10 × index for the standard size and $0.10 × index for the micro, so about $75,000 and $750 of notional at base [3]. The reconstructed change distributions across every team, every calendar day, and six seasons are the table a margin desk would start from.

Read through a SPAN lens, a baseline scan range of about 200 points covers the two-day 99.5% tail on ordinary days (205 up, 187 down) or the one-day game-day 99.5% tail (207 up, 192 down): $2,000 per standard contract, $20 per micro, 2.7% of notional. Stretching to the 99.9% two-day tail gives 300 points and 4%. Either number will look thin beside equity index futures at 5–8% of notional, and that is the point. The 7,500 base is dead notional; margin has to be set on point volatility. From the other side, a market maker's monthly-contract inventory has a P&L standard deviation of $3,720 per contract, so a $2,000 margin is about 0.55 of one monthly sigma — comfortable for a one-day horizon, and a reminder that the leverage is real even though the daily numbers look sleepy. The annual contract's one sigma is $19,600, 26% of notional.

The baseline is wrong on four kinds of day, and each needs an explicit add-on because none is a statistical tail; they are scheduled. Fifty-one team-days in six seasons moved more than 250 points: 17 on the final day of the regular season, when eight ±250 milestones are distributed and a team can take two at once (Montreal and Chicago each printed −500 in April 2023, last in conference plus fewest goals); 9 on a month-end, where roughly one team in five collects a ±125 monthly milestone and ties are paid in full (17–21% of the monthly goals milestones in the sample were ties); 6 on Cup-clinching days, the largest being Vegas at +1,351 on June 13, 2023; and the remaining 19 all playoff games under the 1.5× factor, where the per-game standard deviation is 113 rather than 74. No ordinary regular-season game day appears on the list. An add-on schedule of 125 points on month-end days, 500 on the final regular-season day for teams in contention for a seasonal milestone or the bottom of a conference, and 600 on Final game days for the two finalists would bring the scheduled jumps inside the coverage; without it, a $2,000 baseline is breached by three to six times on those days, and the breach is known a week in advance.

Two portfolio points follow from the correlation structure. A long/short pair in the same game deserves a near-total spread credit for that day, but only for that day; a permanent inter-commodity tier for "team A against team B" would be wrong 95% of the time, so any credit has to be schedule-aware. And because unrelated teams are near-zero correlated, a market maker carrying all 32 names diversifies by roughly √32 at the portfolio level with no credit at all, which is how most of the risk in this product will actually be held.

The contrast with Kalshi is contract design, not a number. A Kalshi binary is fully collateralised: the clearing question is trivial and the leverage is zero. An FSPI standard contract is margined at roughly 2.7% of a notional that mostly does not move, against a monthly sigma of 5% and an annual sigma of 26%, with scheduled jump days the margin model has to know about. That is a real futures product with a real CCP problem — which is what CME wanted, and also why it will never look like a sports bet to a regulator.

The question a market maker would ask is how much of an FSPI position can be laid off in an outcome book. Regressing reconstructed index changes on realised outcomes gives the ceiling: a perfectly priced Kalshi contract delivers the realised outcome, and no better.

At the game level the increment is 123 points higher on a win than on a loss, so the minimum-variance hedge for one standard contract is about $1,230 of game-winner face per game — and it leaves 41 points of residual, most of it the milestone layer. At the monthly level, net wins explain 74% of the monthly change and leave $1,900 per contract-month that no game-winner book can remove; monthly milestones alone are 15% of monthly variance and are not hedgeable anywhere. At the season level a win total gets to 88%, but the binaries Kalshi actually lists for a season — make the playoffs, win the division, the conference, the Cup — do much worse, and a make-playoffs contract on its own explains 59%. For the full annual contract, wins plus playoff wins reach 91% and still leave $5,900 of basis per contract-year.

That is the structural result the first pass reached qualitatively, with numbers attached. FSPI is roughly 70% (game), 74% (month) and 88% (season) a repackaging of outcomes Kalshi already trades, plus a residual of stat production and milestone jumps that is not replicable in any outcome market. From a market maker's chair the residual is the basis risk of running FSPI against a Kalshi or sportsbook hedge. From a regulator's chair it is the part of the product that is not a bet. From FutureSports's chair it is the design margin they have to defend as meaningful rather than as noise.

The Adjacent comparison closes the loop. An NFL Team Index built as an equal-dollar basket of Kalshi and Polymarket outcome contracts is, by construction, 100% replicable in the binaries it is built from. Its basis to Kalshi is zero. That is what would make it a clean settlement reference for a future on prediction-market prices, and it is precisely why it inherits Kalshi's legal status whole: a contract whose only consequence is the P&L of other people's bets has a hard time with the Ninth Circuit's test, which asks whether an event is "inherently associated with a potential financial [or economic] consequence, not just that the event or contingency have some potential downstream financial consequence," and which endorsed rejecting consequences "extrinsic to the parties to the contract" [5]. FSPI's 30% of something else is not an inefficiency. It is the argument.

The index is not a martingale. It drifts upward by 8.8 points per regular-season team-game, about +720 over a season, which is why the mean end-of-season level in the reconstruction is roughly 8,250 rather than 7,500, and why a fairly priced monthly future should sit about 115 points above spot at listing — a testable prediction for the first weeks of trading. The drift has a specific source. Takeaways are credited at +2.5 with no offsetting "opponent takeaway" constituent, while giveaways cost only −0.8, so the takeaway/giveaway term alone contributes +10.6 per game; penalties subtract 4.3; the asymmetric milestones (the +30 clinching goal has no negative twin; the +10 save band is smaller than the −12.5 band) add about 1.6.

Takeaways and giveaways are recorded by the home arena's off-ice crew, and the building effect swamps the team effect. Aggregated over a season the takeaway/giveaway term averages +907 points with a standard deviation of 238 and a range from 362 to 1,658 across team-seasons, and it correlates only 0.42 with the team's total index gain and 0.30 with its goal differential. In contract terms that is a ±$2,400 one-sigma annual component per standard contract determined largely by which building a team plays in and who is in the press box. For a product that will be defended under Core Principle 3 — the same principle Duffy invoked at the CFTC's Innovation Advisory Committee in August against roughly 2,500 self-certified event contracts [10] — the largest deterministic term in the index being an unaudited, rink-dependent judgment call is the first thing a skeptical reviewer will find. It is also fixable without changing the character of the index: net takeaways, or a league-normalised takeaway rate, would remove the drift and most of the building effect in one line of the methodology.