How the MLS model works

Evidence & method

How the forecast is built, and what it was tested against.

Board metadata: 15 upcoming fixtures · 1 model view · training cutoff Oct 17, 10:30 PM EDT · 2 league-wide odds requests, no event calls.

Market boardLatest captured consensus beside the model's fair pricesOPEN

Match	Kickoff	Model lean	Model fair price	Market home	Market draw	Market away	DNB research
D.C. United at FC Cincinnati	Oct 21, 7:00 PM EDT	FC Cincinnati	-149	56.6% · -130	21.5% · +366	22.0% · +355	Not available
Nashville SC at CF Montreal	Oct 24, 1:30 PM EDT	Nashville SC	-123	20.3% · +394	22.0% · +354	57.7% · -137	Not available
Toronto FC at Columbus Crew SC	Oct 24, 3:30 PM EDT	Columbus Crew SC	-138	58.0% · -138	21.5% · +364	20.5% · +388	Not available
Inter Miami CF at New York Red Bulls	Oct 24, 5:30 PM EDT	New York Red Bulls	+151	28.9% · +246	22.0% · +355	49.1% · +104	Not available
Chicago Fire at Atlanta United FC	Oct 24, 7:30 PM EDT	Atlanta United FC	+138	39.6% · +153	25.0% · +300	35.4% · +182	bovada / bovada
FC Cincinnati at D.C. United	Oct 24, 7:30 PM EDT	D.C. United	+147	36.6% · +173	24.5% · +307	38.9% · +157	Not available
New England Revolution at Philadelphia Union	Oct 24, 7:30 PM EDT	Philadelphia Union	-159	59.7% · -148	21.1% · +375	19.3% · +419	Not available
New York City FC at Orlando City SC	Oct 24, 7:30 PM EDT	Orlando City SC	-112	41.4% · +142	24.3% · +311	34.3% · +192	betonlineag / betonlineag
Vancouver Whitecaps FC at Colorado Rapids	Oct 24, 8:30 PM EDT	Colorado Rapids	+115	30.9% · +223	25.3% · +295	43.8% · +128	Not available
Austin FC at Sporting Kansas City	Oct 24, 8:30 PM EDT	Sporting Kansas City	-103	43.2% · +131	24.3% · +311	32.4% · +208	Not available
St. Louis City SC at Real Salt Lake	Oct 24, 8:30 PM EDT	Real Salt Lake	+141	36.4% · +174	25.1% · +299	38.5% · +160	bovada / bovada
San Jose Earthquakes at FC Dallas	Oct 24, 8:30 PM EDT	FC Dallas	-111	52.5% · -110	23.4% · +328	24.1% · +314	Not available
Minnesota United FC at Houston Dynamo	Oct 24, 8:30 PM EDT	Houston Dynamo	+113	47.2% · +112	25.4% · +294	27.4% · +265	Not available
Charlotte FC at Portland Timbers	Oct 24, 10:30 PM EDT	Portland Timbers	+101	48.1% · +108	22.9% · +337	29.0% · +244	Not available
Seattle Sounders FC at San Diego FC	Oct 24, 10:30 PM EDT	San Diego FC	+104	43.8% · +128	24.1% · +314	32.1% · +212	Not available

Compare the model probability to the no-vig market probability. A difference is not automatically an edge.

Model tournamentChronological backtest — why there is no moneyline pickOPEN

Model	Inputs	Moneyline log loss	Moneyline accuracy	Total MAE	Exact score
Poisson · recent xG	Season-transition xG through April; standard xG after	1.0321	49.4%	1.358	11.1%

Why there is no moneyline pickClosing market 1.0145 vs best learned three-way model 1.0257

Lower log loss is better. The market remains the better probability forecast, and every tested model-versus-market edge threshold lost money. “Moneyline” on this page means probabilities only.

Historical tier backtest — not the live record above

Reference tier	Historical games	Avg. confidence	Accuracy	Closing-price ROI
Standard confidence	2,153	50.0%	47.8%	-5.7%
Lower confidence	917	41.8%	39.8%	-2.9%
Higher confidence	1,166	59.8%	59.7%	-3.8%

Across thousands of historical games, higher-confidence forecasts were more accurate and every tier still lost to closing prices. The live table at the top of this page is a separate, far smaller sample and may disagree; where they conflict, neither is settled until the live count grows.

Data auditWhat feeds the forecast, and what remains researchOPEN

Used in the live forecast

Historical match results and game-level expected goals (xG)
5- and 10-match goal, xG, and points form
Rest days, season experience, and regressed Elo
Prior-season points, goal difference, and xG in the March–April route

Available for tracking or research

Closing home/draw/away prices for historical market evaluation
Upcoming fixtures from one league-wide market capture
Live and final MLS scores for the locked reference scorecard
Travel, altitude, time-zone, and home/away form features under evaluation
ASA player and roster-production data under evaluation

Still needed before promotion

Confirmed absences, suspensions, and goalkeeper status
Projected and confirmed lineups plus current continuity
Kickoff weather, heat, wind, and field conditions
Opening and closing totals/moneyline history by book
Current manager status and match-level shot data

How to read this pageScore modes, expected goals, and API-credit policyOPEN

Outcome accuracy: how often the model's most likely result — home, draw, or away — matched the final score. It is compared with the most-common-result baseline, because a model that never calls draws is not honestly described by a one-in-three guessing floor.

Three-way log loss: grades the probability assigned to the result that happened, including draws. Lower is better; an overconfident miss is penalized more heavily.

Confidence gap: what the model claimed minus what it delivered, on the same settled matches. A negative gap means it was more certain than it earned.

95% interval: the range of true accuracies consistent with this sample. Wide intervals mean too few settled matches, not a stable result.

Most likely exact score: the single scoreline with the highest probability, not the average score. A 1–1 modal score can still coexist with a 3.00 average total because the remaining probability is spread across many higher-scoring outcomes.

Moneyline view: home, draw, and away probabilities derived from the score distribution. The American odds are fair, no-vig equivalents—not sportsbook quotes.

Market comparison: each bookmaker's vig is removed before averaging home/draw/away probabilities. The displayed point difference is diagnostic only—not a bet signal or a model input.

Credit policy: one totals-only league request supplies the fixture slate. BTTS and event-specific requests remain off.