His average said over. The shape of his starts said under.

Chris Sale came into NLDS Game 3 averaging 20.4 outs over his last five starts. The line was 18.5. Here is the read behind taking the under, step by step, including the step we did in the wrong order.

Oct 7, 2026 · Dodgers at Braves, Oct 6 · Prices from our 4:31 PM ET board · 10 min read

Step 1: Turn the line into a question

A pitcher-outs line of 18.5 doesn’t ask whether Sale is good, or whether he averages more than 18.5. It asks one yes-or-no question: will he record a 19th out? Eighteen outs is six full innings, so the over needs him to get at least one out in the seventh.

That is why the dashboard average misleads here. Over his last 10 starts Sale averaged 19.2 outs, but only 3 of the 10 reached 19. The average is lifted by a 27-out complete game against these same Dodgers on Aug 27. That game adds almost a full out to the 10-start average, yet it counts as one over, exactly like a 19-out start would. Take it out and the remaining nine starts average 18.3.

When you look at a prop, ask “how often does he clear this line?” and “where do his results bunch up?” before you ask “what does he average?”

Step 2: Find the spike

In this ten-start sample, Sale’s outs pile up at 18: six innings, somewhere between 94 and 108 pitches, then the bullpen. His median start was 18 outs. That is a pattern worth investigating, not proof of a fixed managerial rule or a 60% chance next time.

That pile-up decides which line you want. Different lines treat that most common result differently:

An under at 17.5 is a bet against his habit. An under at 18.5 bets with it, but costs a lot: the books priced it around −200. A whole-number 18 turns his habit into a refund, so the under there only has to beat one thing: him getting an out in the seventh.

Step 3: Get three opinions on the same start

We split the start into three results and asked three sources how likely each was: the sportsbooks (with their margin removed from the 17.5 and 18.5 lines), our model, and Sale’s own last 10 starts.

On the over they roughly agree: 19 or more outs somewhere between 30% and 37%. Where they disagree is the middle. His record says exactly 18 happens most of the time, the books say about 22%, and our model says only 15%.

The model is the outlier, and we know why. It turns its projection into a smooth bell curve around 16.9 outs, so it has no way to know that Atlanta’s manager likes to end Sale’s night exactly at six. That may overstate the appeal of Under 17.5 if the six-inning pattern persists; ten starts alone cannot establish that the model is wrong. It rated Under 17.5 at +130 as an eligible pick worth about +23%. We skipped it, because it loses on Sale’s single most likely result.

Step 4: Know why the model is below his average

The model’s 16.9 sits below every average on the dashboard, by anywhere from 1.2 to 3.5 outs. Before trusting a number like that, find out where the gap comes from.

1. It starts from his recent form, but cautiously. His last five starts produced 102 outs. The model adds two imaginary 15.5-out starts, a typical starter, so five games can’t steer it too hard: (102 + 31) ÷ 7 = 19.0.
2. A machine-learning step cuts it to 16.9. The model learned from about 4,100 starts, with 47 inputs including the opponent, both bullpens, the ballpark and whether it’s a playoff game. Three plausible reasons for a lower forecast are: Pitchers coming off a strong stretch usually come back toward normal. The Dodgers scored the 6th-most runs in baseball (4.77 a game) and struck out less than most lineups (25th of 30), so more balls in play and longer innings. And playoff starts run shorter.
3. The final adjustment barely moves it (16.94).

We can’t split the cut exactly between those reasons; the trained model doesn’t report that. But we can check whether it’s the kind of model that cuts too much in October. Our published validation says the opposite. In a test on the 2025 postseason (87 starter forecasts across 46 games), the model projected 16.2 outs per start and starters actually recorded 14.3. October hooks were shorter than even the model expected. And when it gives an over about a one-in-three chance, it has been about right: forecasts near 33% hit 33% in the regular-season test, and forecasts near 36% hit 38% in the playoff test. Those are small samples, and that test used an earlier version of the model, so we treat this as reassurance, not proof.

Step 5: Shop the line, not just the price

With three sets of chances, every quote on the board becomes simple arithmetic: what does this price need to be true, and how does each source rate it? Here are the board prices as of 4:31 PM ET.

Expected return is the profit per dollar staked if that source’s chances are right; pushes return the stake. “Books” uses the books’ own prices with the margin removed. Break-even for Under 18.5 at −200: 19+ outs must happen less than one time in three.

Two bets stood out. Under 18.5 at FanDuel (−200) was the best under at a U.S.-licensed book, but only slightly favorable by the model and slightly unfavorable by the books’ own prices: close to a coin flip on value. Before first pitch FanDuel moved it to −210, which needs 19+ outs below 32.3% of the time. That left our model with almost nothing (about +1%), so at that price we would pass.

Under 18 at +100 was the cleanest version of the idea, because the push on 18 removes his most common result from the bet. Its other side was −140: with the margin taken out, that book leaned slightly toward 19+ over 17-or-fewer (about 54/46). The other books’ 17.5 and 18.5 prices leaned the other way, about 47/53. That disagreement is a potential pricing edge, conditional on those cross-book probabilities being a fair comparison. The catch: the only book quoting a whole-number 18 was BetOnline, which is offshore and not licensed in U.S. states. Check that any sportsbook you use is licensed where you are.

That was the position we logged in our tracker: Under 18 at +100.

Step 6: Check what the model can’t see, before you bet

This is the step we did in the wrong order. We logged the under at 5:23 PM ET. Our check on Atlanta’s pitching staff came back four minutes later, and almost all of it pointed toward a longer night for Sale:

• Manager Walt Weiss had already shown a long leash. In Wild Card Game 1 he sent Sale out for the seventh, and Sale got 19 outs on 95 pitches.
• The model thought Sale was on a week’s rest. It didn’t know he had thrown 21 pitches in relief to close out the Wild Card Series on Oct 1.
• Atlanta’s pen was short. Didier Fuentes and Ray Kerr, the relievers who can cover several innings, were expected to be unavailable after heavy work in Game 2, and Game 4 was the next day.
• Weiss said he wanted length. The team had lined Sale up for Game 3 to get, in Weiss’s words, a “full-strength Chris Sale.”

Put that together and our judgmental estimate for 19+ outs rose to about 40%, a little above the books. This was a sensitivity check, not a recalibrated model forecast. Holding the books’ estimated chance of exactly 18 at 22.3% leaves about 37.7% for 17 or fewer; under those assumptions, Under 18 at +100 is worth about −2%, not the model’s +22%. We kept the bet, since a six-and-done start would only be a push, and added nothing. Had this check come first, the right call would have been to pass or bet less.

Two things kept it from being a clear mistake. The market had all of this news too, and FanDuel’s under still moved from −200 to −210 toward our side. And the push still covered Sale’s most likely outcome. But the lesson stands: read the bullpen and the manager before you place the bet, not after.

Step 7: Watch the pitch count, not the score

For a live outs prop the number to watch is pitches per inning. In his last 10 starts Sale threw between 94 and 109 pitches, and the hook usually came after six innings. At about 15 pitches an inning, six innings would take 90 pitches, leaving a plausible path to the seventh. At 16 to 17, six innings would take 96 to 102 pitches. These are pace scenarios, not automatic hooks: score, baserunners, matchups and the manager still determine whether he returns.

Sale was at 65 pitches before the fourth inning was over. In that fourth, Kiké Hernández hit a two-run homer and Freddie Freeman added a solo shot. He finished four innings at 80 pitches, 20 an inning, and needed about 94 to get through the fifth. Weiss went to the bullpen, and Sale’s night ended at 15 outs: five innings, seven hits, three runs, seven strikeouts, no walks. Yoshinobu Yamamoto outpitched him, and the Dodgers won 3–1.

That’s why the bet tracker now shows the live pitch count beside the outs for every pitcher prop.

What one result proves (not much)

• The under won for a reason we didn’t predict. Our pregame case was about Sale’s six-inning habit and October hooks. What actually ended his night was a 20-pitch-an-inning pace and two home runs in one inning. That is a good outcome, not evidence for the method.
• Closing-line value can’t judge this bet. Our tracker scored it −7.7% against the closing line. BetOnline was the only book at 18 and never moved its price, so the “close” was simply that book’s own lean toward the over, the same lean we were betting against. Our closing-line check only compares the exact same line, so for a one-book, whole-number line it has nothing to compare against. We count this one as inconclusive.
• The model has two known gaps. A smooth curve misses a manager’s habits, like the spike at 18. And relief outings and reliever availability are not inputs yet. Both mattered for this start.

The checklist

1. Translate the line. 18.5 outs means “an out in the seventh.” Write down the exact result that wins.
2. Look past the average. Check the hit rate over the line and the median; one complete game can move an average by a full out.
3. Find the spike. If a manager habitually pulls a pitcher at the same point, pick the line that treats that result the way you want: win, push, or avoid.
4. Get three opinions. The books with their margin removed, a model, and recent results. Know why they disagree before you side with one.
5. Do the break-even math at each book. A −200 under and a −225 under are different bets; one can be fair and the other not.
6. Check what the model can’t see, first. Bullpen availability, recent relief outings, a manager’s quotes and the next day’s schedule.
7. Set a price you won’t chase. We would take FanDuel at −200 but not at −210.
8. Watch pitches per inning live. Around 15 is the line between “back out for the seventh” and “done after six.”