Most AI predictions you see are a vibe with a logo on it. A number appears, no one tells you where it came from, and you are supposed to trust it. We think that is backwards. So here is the whole engine behind the Moss World Cup Prediction Agent, end to end, in plain language, plus the real Round of 32 match where it priced Morocco over the Netherlands while the market had it backwards. If you want more breakdowns like this on AI, trading, and prediction markets, follow @MossAI_Official, because this is the work we do.
The one-line version
The agent does not guess winners. It prices goals, and then it reads every market off the goal math. The pipeline looks like this:
team strength, turned into a 9-factor index, becomes a pair of expected goals, gets refined by a few deliberate corrections, gets expanded into a full grid of scoreline probabilities by a Dixon-Coles model, and from that grid you can read the probability of every market: win, draw, loss, exact score, over and under, both teams to score, and the handicap. Then each of those is compared to the live market to find where the price drifted.
That last step is the whole point. The model produces an independent probability, and the market price is only used as a reference and a target, never as the answer. Here is each stage.
Step 1: 9-factor strength index
Strength is not one stat, so we do not treat it like one. Every team gets a pre-match strength score built from nine weighted factors, with the weights summing to one. The biggest weight goes to squad quality, the level of the clubs the players actually play for plus the team's ranking. Then recent form with time decay, squad availability for injuries and suspensions, home advantage including altitude and travel, attacking output, defensive solidity, key-player form, a market signal taken from de-vigged bookmaker odds, and head-to-head history.
Squad quality carries the most weight on purpose. In national-team football, individual talent is the most stable signal of strength, while goals and form get distorted by lopsided qualifiers where a strong side runs up the score on a weak one. Leaning on who the players are, not just what the scoreline said last week, is what stops the model from overrating a team that just feasted on minnows.
Step 2: From strength to expected goals
The strength index, combined with each side's attack and defense numbers, produces a pair of raw expected goals, one for each team, with a small home adjustment. Call them lambda.
The single most important thing to understand here is that expected goals is an average, not a scoreline. A lambda of 1.9 does not mean a team scores 1.9 goals. It means that if you replayed this exact match many times, they would average 1.9. Every probability the agent outputs is derived from these two numbers, so getting them right and honest matters more than anything else downstream.
Step 3: The corrections, and the one knob that defines everything
Raw expected goals are not good enough, so they pass through a few deliberate corrections. Two of them only shift strength between the two teams without changing the total goals, which keeps the over-under read stable while sharpening the win-loss read. The corrections blend an Elo-based estimate with the goals-based one to undo qualifier inflation, suppress a weak side's goals when the gap is large so upsets are not overstated, nudge the goal share toward the more talented team without changing the total, and apply small psychological adjustments for must-win games, groups of death, and knockouts.
The core idea is the last one, supremacy widening. Split the two expected goals into two quantities. The total, the sum of both, which we hold fixed. And the supremacy, the difference between them, which measures how lopsided the match is. We then multiply only the supremacy by a widening factor. Widen it and the win, handicap, and both-teams-to-score reads get sharper, while the over-under read does not move at all, because the total was untouched.
That widening factor is the single switch that separates two worlds. For ordinary international matches we set it high, around 1.5, because the strong side usually is just better. For the World Cup we set it almost to one, around 1.05, so we barely widen at all. The reason is not timidity. It is data. On historical World Cup matches, widening to 1.5 makes the model overconfident and increases its error, while staying near 1.05 is the most accurate. Our own backtests put World Cup favorites winning only around half to a little over half of their matches, far below what the public assumes. So the model is deliberately, measurably conservative about favorites at the World Cup, and that is a designed feature, not a bug.
This is why our number for a big favorite is often lower than the bookmakers and the crowd. They tend to be overconfident about star teams at the World Cup. The history says upsets are common. We price for the history.
Step 4: The grid that produces every market
The corrected expected goals go into a Dixon-Coles model, which is the standard Poisson goal model with a correction for the correlation in low scores like 0-0, 1-0, 0-1, and 1-1 that a plain Poisson gets wrong. The output is a full grid, one probability for every scoreline.
Every market is then just a sum over the right cells of that grid. Win, draw, and loss are the cells where the home team scores more, the same, or fewer. Exact score is a single cell. Over and under is the sum of all cells above or below the goals line. Both teams to score is every cell where each side has at least one. The handicap is the sum of cells where the favorite still covers after the line. One coherent grid, every market read consistently off it, no contradictions between them.
Step 5: The edge, model versus market
A probability on its own is just an opinion. The agent turns it into a signal by comparing it to the market. It takes the live odds, strips out the bookmaker or platform margin to recover the true implied probability, and subtracts. When the model's probability is higher than the de-vigged market probability by enough, around four percentage points, the market has underpriced that option and there is value. The agent does this for the win-draw-loss markets and all the main scorelines and surfaces the biggest gap as the match's best value spot.
This is the part that mirrors how the sharp automated wallets actually operate. The edge is never the pick. It is the gap between a calibrated read and a price that drifted.
Two passes: pre-match and lineup-confirmed
The agent reads each match twice. The first pass runs ahead of time on a predicted lineup and gives you the early read. The second pass runs about an hour before kickoff, once the official lineup is published, and recomputes squad quality, expected goals, and every probability on the real eleven. Lineups land roughly an hour out, so that second pass is the sharpest version, and the agent checks upcoming matches every fifteen minutes so it refreshes the moment the team sheet drops.
The real case: Netherlands vs Morocco, called before kickoff
Here is the agent's actual output on a real Round of 32 match, Netherlands vs Morocco, pulled straight from moss.site/wc2026.
On the power index, the agent had the Netherlands ahead, 83.5 to 71.7. So the stronger side overall, clearly. But look at the nine factors underneath that headline, because this is where it gets interesting. The Netherlands led on squad quality (95.4 to 91.7), attacking (99.6 to 81.4), venue (85 to 40), and head-to-head (100 to 0). Morocco led where it mattered most for this specific matchup: recent form (77 to 70.9) and, crucially, defending (80.6 to 63.8). A high-talent, high-attack team that does not defend, against a slightly less starry team that defends hard and is in form. That is a classic upset profile, and the factor breakdown surfaced it instead of burying it under one strength number.
That fed into nearly level expected goals, 1.5 for the Netherlands against 1.6 for Morocco. Run through the grid, the model probabilities came out Netherlands 35.4%, draw 24%, Morocco 40.6%. Now the part that matters. The agent placed each of those next to the live Polymarket price. The market had the Netherlands at 40.3% and Morocco at just 28.4%. The model had Morocco a full 12.2 points higher than the market did, crossed the value threshold, and flagged Morocco win as the recommended value bet at 3.51.
The market saw the famous attacking names and the bigger reputation and priced the Netherlands as favorites. The model saw a team that cannot defend meeting a disciplined side in form, priced Morocco as the likelier winner, and called the market underpriced on Morocco by double digits.
Morocco knocked the Netherlands out. The tie finished level and went to penalties, and Morocco advanced, so the side the model had priced highest and the market had underpriced was the side that went through.
That is the entire design working as intended in one match. Not a vibe, not a lucky pick. A goal model that refused to over-rate the glamour team, a factor breakdown that exposed the defensive mismatch the crowd ignored, and an edge calculation that turned all of it into a single, specific, correct call against the market price. It is the same reason this World Cup has been an upset machine, Spain held by Cape Verde in their opener, Ecuador over Germany, and the model is built to expect exactly that rather than get steamrolled by it.
The takeaway
The agent is not a black box and it is not a vibe. It prices goals with a 9-factor strength read, refines them with corrections that keep the goal total honest, expands them into a full scoreline grid with Dixon-Coles, reads every market off that grid, and compares each one to the de-vigged market to find where the price drifted. And it is tuned, with real backtests, to respect how upset-prone the World Cup actually is, which is why it had Morocco over the Netherlands at a 12-point edge while the market had the Dutch in front. That disagreement was the whole point, and it was right.
Go to moss.site/wc2026 and spend your Moss Diamonds to unlock the full read on any World Cup match: the win, draw, and loss probabilities, the expected goals, the most likely scoreline, the over-under and both-teams-to-score, the recommended and conservative handicap lines, and the model vs live-Polymarket comparison.
Besides, we are also having World Cup special drops for you to earn free Moss Diamonds to unlock the prediction details of the World Cup matches. Stay tuned to our X official account for new events.
