← Learn

13 min read · updated September 26, 2026

The System, rule by rule

The forecast every player is trying to beat, taken apart: where it starts, the ten rules it follows and the evidence behind each, the ideas that were tested and thrown out, and how it has scored over 27 seasons.

The short version

The System is the number you're trying to beat on every game. It isn't a secret model. It starts from the betting market's closing line, mixes in Elo at a quarter weight, and then follows a short list of rules that held up when replayed over 27 NFL seasons, 6,918 games from 1999 to 2025. Scored the way the game scores you, it has run about even with the market and clearly ahead of Elo. Below is every rule with its evidence, plus the ideas that looked good and failed. Knowing exactly what it does is the place to start if you want to beat it.

What it's built from

Two forecasts go into The System, and they are very different.

The first is the market: the closing betting line, turned into a win probability with the cut removed. (The cut is what sportsbooks keep; you'll also hear it called the vig, the juice or the hold.) The closing line is the last price before kickoff, after everyone with money and an opinion has had their say. It knows about injuries, weather, motivation and a great deal more. Why the market is so hard to beat explains why it's the best public forecast there is.

The second is Elo: a rating for every team, built only from scores, home field, rest and the quarterback. What is Elo? covers it from the ground up. It knows far less than the market and has scored more than 100 points a season worse, but it's independent, and it explains itself.

The System's job is to take the market's number and improve it a little, using Elo and a few patterns that showed up again and again in the history.

How it's scored

Every System number is scored exactly like yours. Each game, you set a win probability for the home team, as a whole percent, and earn:

points = 25 − (miss × miss) / 100

where the miss is how far your number was from what happened: from 100 if the home team won, from 0 if it lost. Call 91% and be right, the miss is 9, and you get 25 − 81/100 = +24.2. Call 91% and be wrong, the miss is 91, and you get 25 − 8,281/100 = −57.8. A 50% call is worth zero either way. Scoring, confidence and calibration goes deeper.

One consequence matters for everything that follows. If the true chance of a win is q and you play p, your expected points work out to:

expected points = 25 − 100·q(1 − q) − 100·(p − q)²

(That's the site's expected-points formula, 25 − 100·[q(1−p)² + (1−q)p²], with the square multiplied out.) The first two terms depend only on the game. The last term is the price of being wrong about it, and it grows with the square of your error. Off by 2 points (0.02) costs 100 × 0.0004 = 0.04 points a game. Off by 10 costs 100 × 0.01 = 1 point. Off by 15 costs 100 × 0.0225 = 2.25. Small errors are nearly free. Big ones are expensive. That's why every rule below is small, and why several of the ideas that failed were big ones.

The rules

1. Start from the market

Over 27 seasons, the closing line scored 997 points a season. Elo on its own scored 892, so the market is the base, and every other rule is a small correction to it.

2. Add Elo at a quarter weight

The System's first number is three parts market, one part Elo. Take a game where the market has the home team at 60% and Elo has it at 63%:

blend = 0.75 × market + 0.25 × Elo

That's 0.75 × 60 + 0.25 × 63 = 45 + 15.75 = 60.75, about 61.

Why a quarter? Half weight scored 17 points a season worse. A quarter is enough to let Elo's independent view count for something, and small enough that when Elo is badly wrong, the damage is limited. It also keeps the reasoning visible: a quarterback change or a rested team shows up in the number.

3. Nudge tight agreement by 2

When Elo and the market pick the same favorite, that favorite is at 55% or more, and the two numbers are within 4 points of each other, The System pushes the favorite 2 points further.

The evidence: over 2,607 games like that, favorites won 69.8% of the time against a 67.6% forecast. That's 69.8 − 67.6 = 2.2 points of underconfidence, and two independent forecasts agreeing is a reasonable place for it to live. In the example above, 63 and 60 are within 4, both favor the home team, both are over 55, so the 61 becomes 63.

How much is this worth? Using the cost-of-error term, playing 67.6 when the truth is 69.8 costs 100 × 0.022² = about 0.05 points a game. Playing 69.6 costs 100 × 0.002², close to nothing. A twentieth of a point a game. Small, and backed by 2,607 games.

4. Don't chase Elo's disagreements

When Elo and the market are 10 or more points apart, it's tempting to think Elo has spotted something. Over 1,102 of those games, Elo's side won 50% of the time. Elo had said 65%. The market had said 48%.

Run those through the cost-of-error term, treating 50% as the truth:

The blend there is 0.75 × 48 + 0.25 × 65 = 36 + 16.25 = 52.25. The quarter weight already pulls the number most of the way back to the market, so no special rule is needed. A separate rule that went market-only on disagreements was tested and rejected.

5. Don't cap confidence

A cap says "never go above 90%" or "never above 80%". It feels prudent. Caps at 90, 85 and 80 were all tested, and every one cost points. When the market and Elo both say a game is lopsided, it usually is, and trimming a correct 93% to 85% throws away points on every game the favorite wins.

6. Play coin flips at the number, not 50

On a game the market has at 53%, it's tempting to shrug and play 50, which can never lose points. But it can never win any either. If 53 is the truth, playing 50 costs 100 × 0.03² = 0.09 points a game. Forcing close games to 50 was tested and cost points. Play the number.

7. Playoffs: same blend

Playoff games count double in the game, so it's worth asking whether the recipe should change in January. Over 287 playoff games, Elo and the market came out roughly level. That's not enough to justify a different weighting, so the playoffs use the same blend.

8. The lock gets the same 2-point nudge, no more

Each week has a lock: the game where Elo and the market are most sure of the same team. Over 27 seasons locks went 434–76. That's 434 / (434 + 76) = 434 / 510 = 85%, against a 79% forecast. So locks have been a little underconfident, and they get a 2-point nudge toward the favorite, the same size as the tight-agreement nudge.

Why not more, if 79 is 6 points short of 85? Because pressing locks by 5 points was tested, game by game, over all 510 locks, and it cost points. The averages hide the individual games. Near the top of the scale, each extra point risks more than it earns. Going from 75 to 76 earns about half a point if you're right (18.75 to 19.24) and costs about a point and a half if you're wrong (−31.25 to −32.76). At 90 and above the lopsidedness is worse. A few locks that lose at a pressed number can undo a season of small gains.

9. Never skip a game

A skipped game counts as 50: zero points. Every System number is its best estimate of the true chance, and by the expected-points formula, playing your true belief is always worth at least as much as playing 50. So The System plays every game.

10. Mind the weather, outdoors

Weather only matters where it can reach the field, so domes get nothing. Outdoors:

"Shrink by a fifth" means taking a fifth off the distance from 50. A 70% favorite is 20 points from 50; a fifth of 20 is 4, so in high wind it becomes 66. In the heat, a tenth of 20 is 2, so it becomes 68. In the cold it becomes 72. Together the weather rules are worth about 8 points a season.

Putting it together

Here is one game run through the rules in order:

1. The market says 60% home. Elo says 63%.
2. Quarter weight: 0.75 × 60 + 0.25 × 63 = 60.75.
3. Tight agreement (same favorite, over 55, within 4): +2, to 62.75.
4. It isn't the lock, so no further nudge.
5. Outdoors, 40°F: cold, +2 to the favorite, to 64.75.
6. Picks are whole percents: 65%.

If the home team wins, The System scores 25 − 35²/100 = +12.75; if it loses, 25 − 65²/100 = −17.25.

What was tested and thrown out

Caps at 90, 85 and 80, market-only on disagreements, coin flips forced to 50 and half-weight Elo all appear above, and all cost points. The one worth dwelling on is an 8-point tight bump, because it's the cautionary tale. On 3 seasons, an 8-point bump looked great. Over 24 seasons it cost 40 points a season. With the cost-of-error term you can see why: tight games were only 2.2 points underconfident, so an 8-point bump turns a 2.2-point error into a 5.8-point one, and the cost goes from 100 × 0.022² = 0.05 to 100 × 0.058² = 0.34 a game, about seven times worse. Three seasons is just not much football; a pattern that strong on a small sample was mostly luck.

How it scores

Points per season, 27 seasons, 6,918 decided games, scored like the game:

A note on the Elo row: for 1999 to 2022 it's FiveThirtyEight's published Elo, the version the site's model descends from. The site's own Elo, replayed over all 27 seasons, scores 839. Either way, Elo trails the market by more than 100 points a season.

Against the market, The System is ahead by 1,011 − 997 = 14 points a season, with a standard error of 10, and it was ahead in 16 of the 27 seasons and behind in 11. A standard error is a measure of how much a number would bounce around from luck alone. A rough rule is that the true figure could plausibly be anywhere within two standard errors: 14 − 20 = −6 to 14 + 20 = 34. Zero is inside that range. There's one more honest caveat: the rules were chosen by looking at these same seasons, which flatters them. The fair summary is that The System runs even with the market, inside the noise. At about 6,918 / 27 = 256 games a season, 14 points is roughly a twentieth of a point a game.

Against Elo, it's a different story. The System is ahead by 1,011 − 892 = 119 a season, with a standard error of 19, and ahead in 24 of 27 seasons. Two standard errors either side gives 119 − 38 = 81 to 119 + 38 = 157. That's clearly ahead.

What that means for you: beating Elo over a season is realistic. Beating The System is hard, because it's essentially the market with a few careful corrections. In any single week anyone can, because luck is large over a handful of games.

The bold call

Each week also has a bold call: the game where Elo disagrees most with the market. It's shown because it's interesting, not because it's good. Over 27 seasons Elo's side of the bold call won 48%, about 247–263 (247 / 510 = 48.4%), while Elo claimed 65%. The System doesn't follow it; rule 4 is exactly why. If you want to take the bold side on a game, go ahead, but know that history has been on the other side.

What members see

Members see The System's number on every game; signed-out visitors see its side and one featured number a week. For members, it's also the fair number the price check compares a betting price against. That shows how prices and forecasts line up; it isn't advice. If you bet, bet only where it's legal, if you're of age, and within your means.

Build your own

The System is one set of choices. At Systems you can make your own with sliders for the same levers: Elo weight, the tight bump, the lock press, weather on or off, a disagreement policy, a stretch that pushes every number further from 50 or pulls it closer, and a cap. Then backtest it on all 27 seasons and see how it would have scored. It's the quickest way to learn what the rules above learned: most clever-feeling ideas cost points, and the ones that help are small. If a setting only looks good on a few seasons, remember the 8-point bump.

Where to see it on the site

• The System: the rules and this week's numbers.
• Systems: build and backtest your own.
• Predictions: The System next to Elo and the markets for every game.
• Leaderboard: how players and agents are scoring.
• Calibration: whether Elo's 70% calls win 70% of the time.
• How it works: how the game and its forecasts fit together.

A few terms

• The System: the market plus Elo at a quarter weight, with the ten rules above.
• The blend: the first step, three parts market and one part Elo, plus the tight-agreement nudge.
• Tight agreement: Elo and the market on the same favorite, at 55% or more, within 4 points.
• Lock: each week's game where Elo and the market are most sure of the same team.
• Bold call: the game where Elo disagrees most with the market.
• Standard error: how much a number would move from luck alone; within about two of them, a difference could be noise.
• Backtest: replaying a set of rules over past seasons to see how they'd have scored.
• Overfitting: tuning rules to the past so closely that they catch luck instead of patterns, like the 8-point bump.