Nobody can read 19,220 prices before breakfast. That’s why we model.

Models are central to how we find and evaluate bets. Together with data automation, they let us cover thousands of possibilities and assess opportunities we could never reach by hand. The strongest decisions combine that quantitative foundation with research, context and judgment about the game itself.

Oct 8, 2026 · Opinion from Fourth & Value · 9 min read

At 7:07 a.m. Eastern on Sunday, Oct. 4, our morning data jobs finished. The NFL board held 16,916 sportsbook prices for 14 games. Five NHL games added 1,286 more, and two MLB Division Series games another 1,018.

That is a lot of decisions before anyone has read an injury report. Which players deserve attention? Which lines look attractive? Which book offers the best price? And what would have to happen on the field for a bet to make sense?

Models make this work possible. They turn historical performance, opportunity and matchup data into projections, and where we have fitted probabilities, a way to judge the available price. Automation puts those assessments alongside thousands of sportsbook quotes. Together, they help us find bets we would otherwise miss and give us a concrete basis for evaluating them.

Those numbers matter throughout the decision. So does judgment. A changing role, a coach’s intentions or a player’s condition can alter the outlook in ways a model struggles to capture. Each bet deserves a closer look, and the assessment can change when those details come into view.

How big is the haystack?

The table below shows the scale of a day’s betting across four sports. A distinct bet is one side of one market at one number: Daniel Jones over 31.5 passing attempts, say. A price is that bet at one sportsbook. The same bet at five books is one distinct bet and five prices.

These are undercounts. We carry a deliberately narrow slice of what books offer: mostly main lines rather than full alternate ladders, no live betting, no same-game parlays, no futures, and only the player markets we have a reason to track. A sportsbook’s own menu for the same games is far larger.

Overlapping seasons make the board larger still. A Sunday with football, basketball and hockey could put roughly 9,000 to 10,000 distinct bets in front of us using the figures above. The challenge is finding the worthwhile opportunities within all that choice.

The arithmetic of attention

Give each distinct bet 30 seconds. That’s barely enough to read the line, glance at the player’s last few games and check one injury report. At that pace one NFL Sunday takes almost 48 hours. A full MLB Friday takes more than 21. A seven-game NHL Tuesday takes more than 11, and an ordinary NBA night about 21. The Oct. 4 morning, all three sports together, comes to more than 56 hours. And 30 seconds isn’t analysis. It’s a skim.

Working by hand, it is natural to start with the teams and players you know. Models extend that reach. A backup running back’s carries, a fifth starter’s outs or a third-line winger’s shots can receive the same structured analysis as the stars. We can consider an opportunity because the numbers warrant attention, even if the player was never on our radar.

Models give us the reach to find opportunities and the numbers to evaluate them. Judgment connects those numbers to the game being played.

The model stays central to the decision

A useful model gives us an explicit view of what a player or team is likely to do. Where the model supports a probability at the offered line, we can compare that estimate with the odds and assess the potential value. The projection, its assumptions and its sensitivity all remain relevant as we research the bet.

Automation makes those comparisons practical. It matches identical lines across up to 11 books, pairs over and under prices to remove the sportsbook’s margin, checks quote freshness and finds the best available price. That full MLB Friday came back from just 17 requests to our odds provider. Collecting and organizing the data frees our attention for the decisions that need it.

The model also makes the reasoning easier to examine. If the attraction comes from an expected increase in playing time, an opponent adjustment or a price that differs from the rest of the market, we can investigate that specific assumption and judge how much it supports the bet.

Coverage has limits. On Oct. 4, 8,518 of the 15,992 NFL prop prices had no calibrated model: touchdown-scorer markets, longest-reception lines and players with too little history were among the gaps. Another 680 NFL forecasts were withheld after reliability or sensitivity checks. Keeping those gaps visible helps us distinguish a supported estimate from a number we cannot yet trust.

One Sunday, from haystack to card

Here is what the Oct. 4 morning edition of Today’s Picks did with those 19,220 prices.

1. 19,220Prices on the morning boards6,775 distinct bets on 674 players, across 21 NFL, NHL and MLB games.
2. 546Offers passed a model screen475 NFL, 67 NHL and 4 MLB. Another 680 NFL forecasts were held back as unreliable.
3. 22Got a full reviewDated reporting, injuries, expected opportunity, the price at other books, and a written case and countercase.
4. 9Made the cardSix NFL and three MLB, about one in every 750 of the morning’s distinct bets.

Even with 546 offers passing the model screen, careful review remained a scarce resource. The review stage reached its daily limit in both NFL and NHL, leaving 524 screened offers unreviewed. Ranking helps direct that attention, while the projections and price comparisons provide the foundation for the reviews themselves.

A separate AI-assisted search, conducted without our model’s numbers, received all 21 games to look for opportunity, matchup and participation questions. That is a count of games submitted, not games exhaustively researched. It adds an independent perspective and can bring a researched lead into review even when the model did not flag it. Some mornings, the combined process produces no picks.

Where judgment changes the assessment

Sports are full of details that are difficult to encode consistently: how healthy a player looks after returning, how much a coach trusts a backup, whether a manager is likely to extend a starter, or how a tactical change affects a particular matchup. Some information arrives after the model runs. Some requires interpretation even when we have the facts.

Those intangibles and subjective assessments are a critical part of evaluating an individual bet. They can strengthen the case, reduce our confidence, change our view of the likely outcome or lead us to pass. The useful question is how the detail affects the bet: fewer minutes, a different role, a shorter outing or less opportunity. We can then ask whether the price still makes sense.

A subjective adjustment should have an explanation. We should be able to say what we think the model is missing and how that changes our assessment. Keeping the original forecast visible also lets us learn whether those judgments help over time.

• The market. Prices across up to 11 books provide another estimate to compare with our own. Since Oct. 5, our selection process combines the model probability with the books’ consensus at the exact offered line, giving consensus 75 percent of the weight. That starting weight still needs testing; the blend makes both the model and the market part of the quantitative assessment.
• Current reporting and opportunity. Injury reports, lineups, starting goalies, weather, expected minutes and pitch counts help us judge whether the forecast fits today’s circumstances.
• Independent research. Research conducted without the model’s numbers can surface a factor we missed or a different interpretation of the matchup.
• The rules. Two books can grade the “same” bet differently. See same bet, different rules.
• The price. A good read at a bad number is a bad bet. See the edge was the price.

Two bets, and the assumptions behind them

Take Daniel Jones on Oct. 4. Our NFL model projected 35.3 passing attempts against a line of 31.5. His recent volume blended to 31.6 attempts, and the opponent adjustment lifted the forecast to 35.3. The review marked the bet worth considering and identified the key assumption: “The adjustment, rather than baseline volume, creates the cushion.”

The model had done valuable analytical work: it quantified the matchup’s effect and made the potential opportunity visible. The review could then examine whether that adjustment made sense for the game. Jones finished with 34 attempts. One result cannot establish the quality of a model, but the example shows how a projection and a close reading of its assumptions work together.

Chris Sale’s pitcher-outs prop showed another side of that relationship. His five-start average pointed over the line, while the pattern of his outings and our model supported a closer look at the under. Further research into bullpen availability and the manager’s plans then weakened that case. The model helped evaluate the opportunity; the context changed how much confidence we could place in it. We walk through the full decision in his average said over, the shape of his starts said under.

Better decisions across a bigger board

Models are critical to what we do. They let us evaluate a far larger set of bets, make consistent comparisons and put numbers behind our opinions. Data automation keeps that work manageable. Research and judgment bring in the details that make each game, and each bet, different.

We measure the results because the whole process has to earn our confidence. Our models remain experimental and have not yet been shown to beat betting prices over time. We publish weekly scorecards and research on calibration to examine how the forecasts perform and where they need improvement.

With 19,220 prices on the board, the ambition is to find opportunities we could never cover by hand and understand them well enough to make an informed decision. That takes good models, reliable data and judgment applied one bet at a time.