A live NBA line is a win-probability model's latest output with the sportsbook's margin added on top. It moves after every possession because the score, the clock and the number of possessions left all change after every possession.

That's the difference from the pregame number. Before tip-off a line can sit at the same price for a day. In the November 19 Knicks-Timberwolves game, Covers.com's tracker showed New York opening as a 1.5-point road favorite on November 18 and holding that number until game day. Once the ball goes up, nothing holds. The book's engine recalculates the price as the game state changes, and whether you're betting a live spread or deciding if your pregame points prop still has value at halftime, the number you see comes from that engine.

A model, not a trader

Nobody at DraftKings sets the third-quarter moneyline by hand. Sportsbooks don't publish their live models, and the general betting press agrees that in-game updates come from automated algorithms rather than an oddsmaker typing in new numbers. The engine has to be quick. A technical breakdown from Source Code Lab puts the target refresh cycle for top-tier live feeds at about 300 milliseconds. The same piece says a market should be suspended within 2 seconds of a major event, with an optional 1 to 5 second delay before a bet is accepted while the book checks the price is still current.

Public versions of these models show what goes into them. A guide from AgentBets uses a standard form: win probability equals Φ(margin ÷ (σ × √possessions remaining)), where Φ is the normal distribution, margin is the current lead and σ is the per-possession margin standard deviation, which it puts at 2.0 to 2.2 points. Possessions remaining equal pace × minutes remaining ÷ 48, against a league baseline of about 100 possessions per team per game in 2024-25. A talent gap adds a per-possession drift, and home court is worth 2.5 to 3.0 percentage points before the game starts.

The inputs that move the line follow from that formula:

  • Score margin and clock. These two do most of the work. A lead is worth more the fewer possessions there are left to overturn it.
  • Pace. A fast game leaves more possessions in the same number of minutes, so the same lead is less safe.
  • Foul trouble, injuries, ejections. These change the drift term, because the team on the floor is no longer the one the model was pricing. A big change like this is also why markets get suspended.
  • Runs and timeouts. In a clean model these matter only through the score they produce. In practice they're where the book's price and the model tend to drift apart, as the examples below show.

The late game is a special case. When FiveThirtyEight launched its live NBA playoff win probabilities, it explained that it moved away from the usual logistic-regression approach. It used Poisson scoring rates plus a separate tree-based model for the final minutes, because intentional fouls and timeout strategy make win probability jump in ways a smooth scoring curve doesn't capture.

The juice comes off before you can compare anything

A price on the screen is not a probability until you remove the vig. The conversion formulas are standard. For a favorite, implied probability = |odds| ÷ (|odds| + 100). For an underdog, implied probability = 100 ÷ (odds + 100). Add the two sides together and anything over 100% is the book's margin. Divide each side by that total and you have the fair probability.

Removing the vig from a -150/+130 moneyline

These are the figures from Top End Sports' no-vig walkthrough, worked through in four steps.

-150 implied: 150 ÷ 25060.00%
+130 implied: 100 ÷ 23043.48%
Total implied probability103.48%
Vig (amount over 100%)3.48 points
Fair, -150 side: 60.00 ÷ 103.4857.98%
Fair, +130 side: 43.48 ÷ 103.4842.02%
Back to odds: -(p ÷ (1 − p)) × 100 and +((1 − p) ÷ p) × 100-138 / +138

The book's fair number is close to a 58/42 game. The -150 price asks you to pay as if it were 60/40. A standard -110/-110 market totals 104.76%, a 4.76-point margin, and its fair price is +100 on both sides.

Margins aren't the same everywhere. Mainline moneylines, spreads and totals carry the tightest vig, while props and parlays carry more. Source Code Lab's suggested ranges for operators are 4 to 6% on match result, 5 to 7% on spreads and 12 to 18% on next-scorer-style props. If you price props, the live mainline is the cheapest place to read what the model thinks.

One lead, four prices

This is the live line move you watch every night, shown in numbers. AgentBets' model, for two evenly matched teams with σ at 2.1, gives a 10-point lead these win probabilities depending on the time left. The fair odds column is my own conversion using the formula above, before any vig.

Time left, team up 10 Model win probability Fair moneyline (no vig)
24:00 (halftime) 72% -257
12:00 (start of 4th) 82% -456
6:00 91% -1011
2:00 98.4% -6150

The score didn't change, and the price still went from -257 to -6150. That is the clock running down the possessions remaining, and it explains why a live line keeps moving through a dead stretch of trading empty trips.

The same guide gives two cases where the book's price and the model disagreed. In one, a team led by 7 with about 17.7 possessions left and had a small talent edge. The model said 80.4%, and the sportsbook price implied 73.7%. In the other, a team had just finished a 10-0 run. The model gave the leader 60.9%, while the book implied 65.5%, which the guide reads as the market overreacting to the run. These are one blog's examples, not a study, but they point the same way as the FiveThirtyEight design choice. The places a live model handles worst are the ones that look dramatic: runs, and the last two minutes.

Why DraftKings and Caesars disagree on the same possession

Most books don't build their own price. Unabated describes a follow-the-leader market where a few market makers post the first real number, historically Pinnacle and now Pinnacle alongside BetCRIS, with Circa trying to join them in some markets. Everyone else reprices after the leaders move. Source Code Lab uses the same split for live engines: a follower republishes a data provider's odds with its own margin added, a hybrid adjusts those odds, and a market maker runs its own model.

The delay between those groups can be measured. SportsBookish Research looked at about 12,000 line moves with no news behind them across MLB, NBA and NFL between May 12 and June 1, 2026. Originator-to-follower lag explained about 52% of them. On NBA totals, DraftKings, FanDuel and BetMGM matched an originator's move after a median of about 8 minutes. Caesars, BetRivers, Fanatics and regional books took 25 to 90 minutes. That study covers lines moving without news, not in-play repricing, so treat it as a sign of who follows whom rather than a live latency figure. I couldn't find a public source that measures ESPN BET's or bet365's live repricing speed.

Book Role (Unabated, SportsBookish) Median lag on NBA totals Lakers ML vs Sacramento, Nov 20 snapshot
Pinnacle Originator Sets the move Not listed
Circa Originator (in some markets) Sets the move Not listed
DraftKings Fast follower About 8 min -1040
FanDuel Fast follower About 8 min -1100
BetMGM Fast follower About 8 min Not listed
Caesars Slow follower 25 to 90 min -1000
Fanatics Slow follower 25 to 90 min -1000
theScore Bet Not classified Not measured -1400

That Lakers snapshot from Covers.com shows how wide the gap can get. At -1000 the favorite implies 90.91%, and at -1400 it implies 93.33%, which is more than two points of probability on the same team at the same moment, with no news behind it. VegasInsider's board for Warriors-Grizzlies showed the same thing on the spread: Caesars at -5.5 (-115), Fanatics at -6.5 (-120), Kalshi at -8.5 (-104).

Not a hand-set number, and not the app's win meter

Two things get mistaken for the live line.

The first is the idea that a live price is a forecast. It isn't only a forecast. As Covers.com puts it, oddsmakers' "primary goal is to reduce risk for the sportsbook rather than to paint an accurate picture of reality." The best argument on the other side is that a book that predicts badly gets picked off, so prediction and risk management end up being the same job. That's partly true, and it's why the model sits underneath everything. But books also shade prices to balance their exposure. SportsBookish attributes about 12% of news-less moves to exposure rebalancing, and Unabated notes that Pinnacle has historically offered a sharper price on the side it wants bet. The price is the model's number plus a business decision.

The second is the live win percentage in a betting app. Action Network's Bet Win Probability, built by data scientist Colin Davy, updates as the game state changes, but it tells you whether a ticket you already hold is likely to cash. It is not a price anyone will let you bet at.

Where the windows open

Once you know the line is a model's output, the soft spots become predictable, and that should change what you check before betting live.

Strip the vig before you compare anything, so that -150 means 58% to you, not 60%. Expect the mispricings in the spots models handle poorly: right after a run, and in the last couple of minutes of a close game. When a game state changes, compare a fast follower against a slow one, because the books that update late are the ones still showing the old price. SportsBookish estimates the edge from betting stale follower prices at 0.5% to 1.5% per bet, which is small, real and easy to lose if you don't track it. Logging the price you took against where it closed, the habit behind closing line value, is the only way to tell whether you're catching those windows or just betting more often. If you're still deciding whether in-game betting suits you at all, the trade-offs between live and pre-game betting are covered separately.

Already pricing NBA props before tip-off?

The pillar guide shows how to carry those projections into the live market and catch the windows this model leaves open.

Read the live NBA props guide