{"article":{"slug":"runs-scored-in-world-baseball-classic","title":"Runs scored in World Baseball Classic","subtitle":null,"summary":"kqr works through a Metaculus Cup question on total runs in the 2026 World Baseball Classic, showing why per-game run counts are overdispersed, how to model them with a negative binomial distribution, and how sums of negative binomial counts behave when forecasting a tournament total.","content_type":"blog_post","language":"en","canonical_url":"https://entropicthoughts.com/sums-of-overdispersed-counts","author":{"name":"kqr","url":null,"person_slug":null,"person_url":null},"authored_by":"human","publisher":{"name":"Entropic Thoughts","url":null,"listing_slug":null,"listing":null},"topics":[{"name":"statistics","slug":"statistics","url":"https://listedarticles.com/topics/statistics"},{"name":"forecasting","slug":"forecasting","url":"https://listedarticles.com/topics/forecasting"},{"name":"probability","slug":"probability","url":"https://listedarticles.com/topics/probability"}],"about_listings":[],"cover_image_url":null,"license":"all-rights-reserved","word_count":2440,"reading_minutes":11,"published_at":"2026-11-02T23:00:00.000Z","added_at":"2026-11-04T08:08:32.830Z","updated_at":"2026-11-04T08:08:32.830Z","added_via":"api","contributor":{"type":"agent","name":"ListedStartups Using Bot","registered":true},"profile_url":"https://listedarticles.com/articles/runs-scored-in-world-baseball-classic","markdown_url":"https://listedarticles.com/articles/runs-scored-in-world-baseball-classic.md","example":false,"citation":"kqr, Entropic Thoughts. \"Runs scored in World Baseball Classic.\" 2 Nov 2026. https://entropicthoughts.com/sums-of-overdispersed-counts (all-rights-reserved)","access":{"human_view":"preview","full_text_available":true,"source_url":"https://entropicthoughts.com/sums-of-overdispersed-counts"},"body_markdown":"\nby kqr,\npublished 2026-11-03\n\nTags:\n\n- [statistics](https://entropicthoughts.com/tags.html#statistics)\n- [forecasting](https://entropicthoughts.com/tags.html#forecasting)\n\n![sums-of-overdispersed-counts.jpg](https://entropicthoughts.com/image/banner/sums-of-overdispersed-counts.jpg)\n\nOn the go? [Listen to an ai-narrated version of this article](https://entropicthoughts.com/audio/sums-of-overdispersed-counts.opus).\n\nConfused? [More about ai narration](https://entropicthoughts.com/ai-narration).\n\nIn the [Spring 2026 edition](https://www.metaculus.com/tournament/metaculus-cup-spring-2026/) of the Metaculus Cup one question concerned [the\nnumber of runs in the knockout stage](https://www.metaculus.com/questions/41684/total-runs-in-knockout-stage-of-2026-wbc/) of the World Baseball Classic (wbc)\nin 2026.11 I only got around to finishing this article now, sorry.\n\nThe community predicted a mean of just under 70 runs, with a spread of 60–80. My\nprediction was 60 runs, with a spread of 42–80. The actual number ended up being\n59 runs – nearly exactly what I predicted. Let’s see how that was done.\n\nFirst, there’s a sloppy way of making a quick guess. It’s not very interesting,\nbut it’s a good baseline. Once we have that, we’ll dig into modeling counts on a\nmore granular level, and that’s when it gets interesting. Those techniques can\nbe used for all sorts of counts, not just sports prediction!\n\n# [Baseline forecast](#baseline-forecast)\n\nI don’t know anything about baseball. I don’t even know what a run is, so I had\nmy research assistant count how many runs were in a few of the previous wbcs.\n\n| Year | Runs |\n| --- | --- |\n| 2009 | 33 |\n| 2013 | 12 |\n| 2017 | 18 |\n| 2023 | 76 |\n\nThe numbers from before 2023 are lower because those years didn’t include\nquarter finals in the knockout stages. Thus we can approach a basic model:\n\n| Year | Runs | Games | Per game |\n| --- | --- | --- | --- |\n| 2009 | 33 | 3 | 11 |\n| 2013 | 12 | 3 | 4 |\n| 2017 | 18 | 3 | 6 |\n| 2023 | 76 | 7 | 11 |\n| Σ | 139 | 16 | 8.7 |\n\nSince there will be seven games in 2026, the expectation for the number of runs\nin 2026 is 61.\n\nTo get the uncertainty around the mean, we’ll need the standard deviation of the\nper-game values22 Suggestion: use [the spc hack](https://entropicthoughts.com/estimating-standard-deviation-from-timeseries-plot) to compute this in your\nhead., and it’s somewhere around 4. The standard deviation of a sum of seven\nvariables is the square root of seven times the standard deviation, so around 10\nruns. That makes for an initial symmetric forecast of 43–79 runs, with 90 %\nconfidence.\n\nExcept not so fast. That forecast assumes the thin tails of a normal\ndistribution where we know the exact value of the standard deviation. We don’t\nknow the standard deviation so accurately, because we based our estimation on\nonly four numbers! To account for this, we’ll use [a handy table of correction\nfactors](https://entropicthoughts.com/ninety-percent-of-the-t-distribution). That results in a more realistic forecast of 35–87 runs, with 90 %\nconfidence.33 Hey, this is already a reasonable forecast for the actual\noutcome. I wonder where the community got 70 runs from.\n\n# [Finding underlying structure](#finding-underlying-structure)\n\nBut I think we can do better. If we expand the 2023 count, we see that it can be\nsplit into stages, games, and then teams:\n\n| Stage | Game | Team | Runs |\n| --- | --- | --- | --- |\n| Quarter finals | Japan vs. Italy | Japan | 9 |\n| \" | \" | Italy | 3 |\n| \" | Mexico vs. Puerto Rico | Mexico | 5 |\n| \" | \" | Puerto Rico | 4 |\n| … | … | … | … |\n| Semi finals | Japan vs. Mexico | Japan | 6 |\n| \" | \" | Mexico | 5 |\n| … | … | … | … |\n| Σ | 7 | 14 | 76 |\n\nThis additional structure might let us narrow down our range a little. If we\ntabulate this for all teams in all games in all the four years we had data for,\nwe can summarise the run counts with\n\n| Mean | Variance |\n| --- | --- |\n| 4.3 | 10 |\n\nHere something odd stands out: these are not Poisson events! If they were\nPoisson events, the variance would be roughly equal to the mean, but now it’s\nover twice as large.44 Which is a significant difference, as can be revealed\nby [a chi-square test](https://entropicthoughts.com/chi-square-from-fundamentals). This says runs are not independently spread out, as we\nassumed in the previous prediction, but instead they occur in clumps.\n\nWhen a count comes from a set of events that are not independent, but rather\nseem to encourage each other, we say that the count is *overdispersed*. The\nnegative binomial distribution is a popular choice for modeling overdispersed counts.\n\n# [Interlude on the negative binomial distribution](#interlude-on-the-negative-binomial-distribution)\n\nThere at least five different parametrisations of the negative binomial\ndistribution, but for the purpose of modeling overdispersed counts, the most\nlogical is using \\(\\mu\\) and \\(\\theta\\) to describe the shape of the distribution.\nHere, \\(\\mu\\) is the average count, and \\(\\theta\\) represents the amount of\noverdispersion, i.e. how much more lumpy the count is compared to a Poisson count.\n\nThis is where things come together beautifully. We’ll ignore \\(\\mu\\) for a moment,\nand look at some points of reference for the \\(\\theta\\) parameter:\n\n- Zero: this means no overdispersion, and the distribution is equal to the\n  Poisson distribution.\n- Between zero and one we get a blend of the Poisson distribution and the\n  geometric distribution. The variance goes up and the tail gets thicker.\n- One: this makes the distribution equal to the geometric distribution. At this\n  point, the tail is exponential.\n- Greater than one: here we get a power law with a subexponential tail.\n\nThe plot below renders density curves with a \\(\\mu\\) of 25, starting at a\ntransparent Poisson distribution (\\(\\theta=0\\)) and going up to the solid\ngeometric distribution (\\(\\theta=1\\)). If we trace the curves from most\ntransparent to solid, we see how the distribution walks over to the left, merges\nwith the axis (making lower counts more likely), and morphs from Poisson to\ngeometric, all while extreme counts (far to the right) also become larger in\nmagnitude.\n\n![overdispersed-counts-01.svg](https://entropicthoughts.com/image/overdispersed-counts-01.svg)\n\nIn the plot above we stopped at \\(\\theta=1\\), but we could go further into power\nlaw territory, and it quickly gets crazy.\n\n![overdispersed-counts-02.svg](https://entropicthoughts.com/image/overdispersed-counts-02.svg)\n\nWe won’t use this heavy-tailed region in this article, I just wanted to make\nclear that it doesn’t stop at a \\(\\theta\\) of one – it goes on.\n\n## [Amazing intuition nobody cares about](#amazing-intuition-nobody-cares-about)\n\nLet’s think for a moment about the range between Poisson and geometric, and what\nhappens as we go from one to the other.\n\nImagine flipping a coin. The geometric distribution models the length of the\nfirst uninterrupted sequence of tails. It could be zero if the first toss lands\non heads. It would typically be one or two. But it could also be really long, if\nheads takes its sweet time to show up. In fact, after a sequence of \\(n\\) tails,\nthe probability of the next tails is still 50/50. This makes the variance of the\ngeometric distribution quite high.\n\nThe negative binomial distribution (with \\(\\theta\\) between zero and one) models\nthe total length of *multiple* sequences of tails. Specifically, it models the\nlength of the first \\(r=\\frac{1}{\\theta}\\) sequences of tails. Thus when\n\\(\\theta=1\\) it models the length of the first sequence, which makes it equivalent\nto the geometric distribution. When \\(\\theta=0.5\\), it models the combined length\nof the first *two* sequences, etc.\n\nTo make this make sense, we have to consider what happens to the hypothetical\ncoin as we increase \\(r\\) (decrease \\(\\theta\\)). We meddle with the probability of\nthe coin so that the expected total number of tails in the sequences stays the\nsame. Thus, when we increase \\(r\\), we also increase the probability, so that\nheads are more likely, and thus the length of individual sequences of tails\ndecrease, but their combined length stays the same. Everything else we are going\nto say works even if we don’t do that, but by keeping the expected total number\nof tails the same, we can see more clearly how the variance increases, as in the\nabove plots.\n\nWhen we get into lower \\(\\theta\\) (or higher \\(r\\)), we are now counting the length of a\nlarge number of sequences of tails. The variation in their lengths will sort of\ncancel out and converge to something more stable. Thus the variance of the\nnegative binomial distribution decreases as we decrease \\(\\theta\\) and make it\nless overdispersed.\n\nAt some point we get to \\(\\theta \\rightarrow 0\\), which corresponds to \\(r\n\\rightarrow \\infty\\). At this point, the probability of heads \\(p \\rightarrow 1\\)\nand we have the Poisson distribution. But note what this means for our\ninterpretation of the Poisson distribution: the events we are counting are the\ntotal number of tails in an infinite number of sequences of tails when heads is\nalmost guaranteed to happen on every toss!\n\n*Isn’t all this just nuts?*55 And what about the heavy-tailed region of the\nnegative binomial distribution? That’s when heads get increasingly *rare* but we\nonly count a partial sequence of tails to compensate.\n\n# [Modeling baseball runs as overdispersed counts](#modeling-baseball-runs-as-overdispersed-counts)\n\nGetting back to the problem, we ask our computer to fit a negative binomial\ndistribution to the counts of runs each team scored. It tells us\n\n\\[\\mu = 4.3,\\;\\;\\;r=3.4\\]\n\nWe know from the amazing intuition that an \\(r\\) of 3.4 corresponds to a \\(\\theta\\)\nof 0.30. It’s difficult to tell from \\(\\theta\\) alone how much overdispersion\nthere is66 We actually get a better idea from the variance which was more than\ntwice the mean. That’s why we started suspecting overdispersion in the first\nplace., but it corresponds roughly to the middle curve on the first plot of\nincreasing \\(\\theta\\). If we try to apply the amazing intuition further,\nwe may think of this as tossing a coin and counting the total number of tails in\nthe first 3.4 sequences of tails, when the coin is only slightly biased toward\ntails.77 We can tell the bias because \\(r \\lt \\mu\\) means the coin is biased\ntoward tails, while \\(r \\gt \\mu\\) means the coin is biased toward heads.\n\nI don’t know enough about baseball to translate this into physical processes,\nbut there must be some sort of decision point in the game that happens on\naverage 7.7 times per game, and 44 % of the time when it comes up, it triggers\nan event that stops the opposing team from scoring a run. When the event has\nbeen triggered 3.4 times, the game ends. Note that this doesn’t mean the 3.4th\nevent actually *causes* the game to end, it means only that the game ends on\naverage after 3.4 such events. Similarly, the lack of an event might not\ndirectly cause a run to be scored, it might just be that runs are correlated\nwith the absence of such an event.\n\nAnyway, all of this is irrelevant to forecasting on the question.\n\n# [Total number of counts for 7 games](#total-number-of-counts-for-7-games)\n\nWe modeled the count for one team in one game as \\(\\mathrm{NB}(\\mu=4.3, r=3.4)\\).\nWe want the total number for all seven games in the knockout stages; that’s 14\ndraws from the distribution above.\n\nNow, the negative binomial distribution has a convenient identity:\n\n\\[x\\_i \\sim \\mathrm{NB}(r, p) \\Longrightarrow \\sum\\_i^n x\\_i \\sim \\mathrm{NB}(nr, p)\\]\n\nThis says that *if* we have parametrised the negative binomial distribution with\n\\((r, p)\\) and we want the distribution of the sum of 14 such random values, we\ncan use the negative binomial distribution with parameters \\((14r, p)\\).\n\nWe don’t have \\((r, p)\\), though, we have \\((\\mu, r)\\). We could convert that to\n\\((r, p)\\) and then convert back, but there’s a more fun way: we can prove a\ncorresponding version of the identity for our parametrisation on our own! We\nwant to find \\(\\mu'\\) and \\(r'\\) such that\n\n\\[x\\_i \\sim \\mathrm{NB}(\\mu, r) \\Longrightarrow \\sum\\_i^n x\\_i \\sim \\mathrm{NB}(\\mu', r')\\]\n\nWe start with the parameter transformation\n\n\\[p(\\mu, r) = \\frac{r}{r + \\mu}\\]\n\nwhich lets us rephrase the first identity to\n\n\\[x\\_i \\sim \\mathrm{NB}(r, p(\\mu, r)) \\Longrightarrow \\sum\\_i^n x\\_i \\sim \\mathrm{NB}(nr, p(\\mu', nr))\\]\n\nFor this to be true still, we need to find the value of \\(\\mu'\\) that makes\n\\(p(\\mu', nr) = p(\\mu, r)\\). That means we need to make it so that\n\n\\[\\frac{r}{r + \\mu} = \\frac{nr}{nr + \\mu'}\\]\n\nThis happens when \\(\\mu' = n\\mu\\) because then we get\n\n\\[\\frac{nr}{nr + \\mu'} = \\frac{nr}{nr + n\\mu} = \\frac{nr}{n(r + \\mu)} = \\frac{r}{r + \\mu}\\]\n\nThus, the identity we are looking for is simply\n\n\\[x\\_i \\sim \\mathrm{NB}(\\mu, r) \\Longrightarrow \\sum\\_i^n x\\_i \\sim \\mathrm{NB}(n\\mu, nr)\\]\n\nHow convenient.88 The convenience makes it even more strange that I couldn’t\nfind anyone stating this version of the identity. I must have overlooked\nsomething.\n\nThis means the number of runs scored in the knockout stages of the 2026 wbc is\nprobably distributed as \\(\\mathrm{NB}(\\mu=14 \\times 4.3, r=14 \\times 3.4)\\). That\ngives us a forecast of 42–80 at 90 % confidence.\n\nThe new interval of 42–80 based on more detailed modeling is better (narrower)\nthan the baseline we had at 35–87. That’s what we hoped for. Now that we\napproached it as the sum of 14 independent-but-lumpy counts, we both used more\ndata to estimate the variation, and we average out some of the noise when we sum\nit all together.\n\nGo forth and use the negative binomial distribution when appropriate!\n\n# [Appendix A: On the independence of run counts between teams](#appendix-a--on-the-independence-of-run-counts-between-teams)\n\nIn everything above, we treated the run counts from both teams in a game as\nindependent of each other. I could hear an argument that maybe the run counts\nfor the two teams in the same game are not independent. It would make sense\nfor them to be negatively correlated: a good team scores more runs *and*\nprevents their opponent from doing the same. That would be bad for us, because\nthat would increase the variation in the count, making our forecast overconfident.\n\nFortunately, if anything, the correlation appears to be positive. Though it is\nso weak that fitting the negative binomial model to the number of runs per\ngame rather than per team gives the same forecast interval at 90 % confidence.\n\n# Sidenotes\n\n[1](#fnr.1)\n\nI only got around to finishing this article now, sorry.\n\n[2](#fnr.2)\n\nSuggestion: use [the spc hack](https://entropicthoughts.com/estimating-standard-deviation-from-timeseries-plot) to compute this in your\nhead.\n\n[3](#fnr.3)\n\nHey, this is already a reasonable forecast for the actual\noutcome. I wonder where the community got 70 runs from.\n\n[4](#fnr.4)\n\nWhich is a significant difference, as can be revealed\nby [a chi-square test](https://entropicthoughts.com/chi-square-from-fundamentals).\n\n[5](#fnr.5)\n\nAnd what about the heavy-tailed region of the\nnegative binomial distribution? That’s when heads get increasingly *rare* but we\nonly count a partial sequence of tails to compensate.\n\n[6](#fnr.6)\n\nWe actually get a better idea from the variance which was more than\ntwice the mean. That’s why we started suspecting overdispersion in the first\nplace.\n\n[7](#fnr.7)\n\nWe can tell the bias because \\(r \\lt \\mu\\) means the coin is biased\ntoward tails, while \\(r \\gt \\mu\\) means the coin is biased toward heads.\n\n[8](#fnr.8)\n\nThe convenience makes it even more strange that I couldn’t\nfind anyone stating this version of the identity. I must have overlooked\nsomething.\n","body_html":"<p>by kqr,\npublished 2026-11-03</p>\n<p>Tags:</p>\n<ul><li><a href=\"https://entropicthoughts.com/tags.html#statistics\" rel=\"nofollow ugc noopener\">statistics</a></li><li><a href=\"https://entropicthoughts.com/tags.html#forecasting\" rel=\"nofollow ugc noopener\">forecasting</a></li></ul>\n<figure><img src=\"https://entropicthoughts.com/image/banner/sums-of-overdispersed-counts.jpg\" alt=\"sums-of-overdispersed-counts.jpg\" loading=\"lazy\" decoding=\"async\" referrerpolicy=\"no-referrer\" /></figure>\n<p>On the go? <a href=\"https://entropicthoughts.com/audio/sums-of-overdispersed-counts.opus\" rel=\"nofollow ugc noopener\">Listen to an ai-narrated version of this article</a>.</p>\n<p>Confused? <a href=\"https://entropicthoughts.com/ai-narration\" rel=\"nofollow ugc noopener\">More about ai narration</a>.</p>\n<p>In the <a href=\"https://www.metaculus.com/tournament/metaculus-cup-spring-2026/\" rel=\"nofollow ugc noopener\">Spring 2026 edition</a> of the Metaculus Cup one question concerned <a href=\"https://www.metaculus.com/questions/41684/total-runs-in-knockout-stage-of-2026-wbc/\" rel=\"nofollow ugc noopener\">the\nnumber of runs in the knockout stage</a> of the World Baseball Classic (wbc)\nin 2026.11 I only got around to finishing this article now, sorry.</p>\n<p>The community predicted a mean of just under 70 runs, with a spread of 60–80. My\nprediction was 60 runs, with a spread of 42–80. The actual number ended up being\n59 runs – nearly exactly what I predicted. Let’s see how that was done.</p>\n<p>First, there’s a sloppy way of making a quick guess. It’s not very interesting,\nbut it’s a good baseline. Once we have that, we’ll dig into modeling counts on a\nmore granular level, and that’s when it gets interesting. Those techniques can\nbe used for all sorts of counts, not just sports prediction!</p>\n<h1 id=\"baseline-forecast\"><a href=\"#baseline-forecast\">Baseline forecast</a></h1>\n<p>I don’t know anything about baseball. I don’t even know what a run is, so I had\nmy research assistant count how many runs were in a few of the previous wbcs.</p>\n<div class=\"table-wrap\"><table><thead><tr><th>Year</th><th>Runs</th></tr></thead><tbody><tr><td>2009</td><td>33</td></tr><tr><td>2013</td><td>12</td></tr><tr><td>2017</td><td>18</td></tr><tr><td>2023</td><td>76</td></tr></tbody></table></div>\n<p>The numbers from before 2023 are lower because those years didn’t include\nquarter finals in the knockout stages. Thus we can approach a basic model:</p>\n<div class=\"table-wrap\"><table><thead><tr><th>Year</th><th>Runs</th><th>Games</th><th>Per game</th></tr></thead><tbody><tr><td>2009</td><td>33</td><td>3</td><td>11</td></tr><tr><td>2013</td><td>12</td><td>3</td><td>4</td></tr><tr><td>2017</td><td>18</td><td>3</td><td>6</td></tr><tr><td>2023</td><td>76</td><td>7</td><td>11</td></tr><tr><td>Σ</td><td>139</td><td>16</td><td>8.7</td></tr></tbody></table></div>\n<p>Since there will be seven games in 2026, the expectation for the number of runs\nin 2026 is 61.</p>\n<p>To get the uncertainty around the mean, we’ll need the standard deviation of the\nper-game values22 Suggestion: use <a href=\"https://entropicthoughts.com/estimating-standard-deviation-from-timeseries-plot\" rel=\"nofollow ugc noopener\">the spc hack</a> to compute this in your\nhead., and it’s somewhere around 4. The standard deviation of a sum of seven\nvariables is the square root of seven times the standard deviation, so around 10\nruns. That makes for an initial symmetric forecast of 43–79 runs, with 90 %\nconfidence.</p>\n<p>Except not so fast. That forecast assumes the thin tails of a normal\ndistribution where we know the exact value of the standard deviation. We don’t\nknow the standard deviation so accurately, because we based our estimation on\nonly four numbers! To account for this, we’ll use <a href=\"https://entropicthoughts.com/ninety-percent-of-the-t-distribution\" rel=\"nofollow ugc noopener\">a handy table of correction\nfactors</a>. That results in a more realistic forecast of 35–87 runs, with 90 %\nconfidence.33 Hey, this is already a reasonable forecast for the actual\noutcome. I wonder where the community got 70 runs from.</p>\n<h1 id=\"finding-underlying-structure\"><a href=\"#finding-underlying-structure\">Finding underlying structure</a></h1>\n<p>But I think we can do better. If we expand the 2023 count, we see that it can be\nsplit into stages, games, and then teams:</p>\n<div class=\"table-wrap\"><table><thead><tr><th>Stage</th><th>Game</th><th>Team</th><th>Runs</th></tr></thead><tbody><tr><td>Quarter finals</td><td>Japan vs. Italy</td><td>Japan</td><td>9</td></tr><tr><td>&quot;</td><td>&quot;</td><td>Italy</td><td>3</td></tr><tr><td>&quot;</td><td>Mexico vs. Puerto Rico</td><td>Mexico</td><td>5</td></tr><tr><td>&quot;</td><td>&quot;</td><td>Puerto Rico</td><td>4</td></tr><tr><td>…</td><td>…</td><td>…</td><td>…</td></tr><tr><td>Semi finals</td><td>Japan vs. Mexico</td><td>Japan</td><td>6</td></tr><tr><td>&quot;</td><td>&quot;</td><td>Mexico</td><td>5</td></tr><tr><td>…</td><td>…</td><td>…</td><td>…</td></tr><tr><td>Σ</td><td>7</td><td>14</td><td>76</td></tr></tbody></table></div>\n<p>This additional structure might let us narrow down our range a little. If we\ntabulate this for all teams in all games in all the four years we had data for,\nwe can summarise the run counts with</p>\n<div class=\"table-wrap\"><table><thead><tr><th>Mean</th><th>Variance</th></tr></thead><tbody><tr><td>4.3</td><td>10</td></tr></tbody></table></div>\n<p>Here something odd stands out: these are not Poisson events! If they were\nPoisson events, the variance would be roughly equal to the mean, but now it’s\nover twice as large.44 Which is a significant difference, as can be revealed\nby <a href=\"https://entropicthoughts.com/chi-square-from-fundamentals\" rel=\"nofollow ugc noopener\">a chi-square test</a>. This says runs are not independently spread out, as we\nassumed in the previous prediction, but instead they occur in clumps.</p>\n<p>When a count comes from a set of events that are not independent, but rather\nseem to encourage each other, we say that the count is <em>overdispersed</em>. The\nnegative binomial distribution is a popular choice for modeling overdispersed counts.</p>\n<h1 id=\"interlude-on-the-negative-binomial-distribution\"><a href=\"#interlude-on-the-negative-binomial-distribution\">Interlude on the negative binomial distribution</a></h1>\n<p>There at least five different parametrisations of the negative binomial\ndistribution, but for the purpose of modeling overdispersed counts, the most\nlogical is using (\\mu) and (\\theta) to describe the shape of the distribution.\nHere, (\\mu) is the average count, and (\\theta) represents the amount of\noverdispersion, i.e. how much more lumpy the count is compared to a Poisson count.</p>\n<p>This is where things come together beautifully. We’ll ignore (\\mu) for a moment,\nand look at some points of reference for the (\\theta) parameter:</p>\n<ul><li><p>Zero: this means no overdispersion, and the distribution is equal to the</p><p>Poisson distribution.</p></li><li><p>Between zero and one we get a blend of the Poisson distribution and the</p><p>geometric distribution. The variance goes up and the tail gets thicker.</p></li><li><p>One: this makes the distribution equal to the geometric distribution. At this</p><p>point, the tail is exponential.</p></li><li>Greater than one: here we get a power law with a subexponential tail.</li></ul>\n<p>The plot below renders density curves with a (\\mu) of 25, starting at a\ntransparent Poisson distribution ((\\theta=0)) and going up to the solid\ngeometric distribution ((\\theta=1)). If we trace the curves from most\ntransparent to solid, we see how the distribution walks over to the left, merges\nwith the axis (making lower counts more likely), and morphs from Poisson to\ngeometric, all while extreme counts (far to the right) also become larger in\nmagnitude.</p>\n<figure><img src=\"https://entropicthoughts.com/image/overdispersed-counts-01.svg\" alt=\"overdispersed-counts-01.svg\" loading=\"lazy\" decoding=\"async\" referrerpolicy=\"no-referrer\" /></figure>\n<p>In the plot above we stopped at (\\theta=1), but we could go further into power\nlaw territory, and it quickly gets crazy.</p>\n<figure><img src=\"https://entropicthoughts.com/image/overdispersed-counts-02.svg\" alt=\"overdispersed-counts-02.svg\" loading=\"lazy\" decoding=\"async\" referrerpolicy=\"no-referrer\" /></figure>\n<p>We won’t use this heavy-tailed region in this article, I just wanted to make\nclear that it doesn’t stop at a (\\theta) of one – it goes on.</p>\n<h2 id=\"amazing-intuition-nobody-cares-about\"><a href=\"#amazing-intuition-nobody-cares-about\">Amazing intuition nobody cares about</a></h2>\n<p>Let’s think for a moment about the range between Poisson and geometric, and what\nhappens as we go from one to the other.</p>\n<p>Imagine flipping a coin. The geometric distribution models the length of the\nfirst uninterrupted sequence of tails. It could be zero if the first toss lands\non heads. It would typically be one or two. But it could also be really long, if\nheads takes its sweet time to show up. In fact, after a sequence of (n) tails,\nthe probability of the next tails is still 50/50. This makes the variance of the\ngeometric distribution quite high.</p>\n<p>The negative binomial distribution (with (\\theta) between zero and one) models\nthe total length of <em>multiple</em> sequences of tails. Specifically, it models the\nlength of the first (r=\\frac{1}{\\theta}) sequences of tails. Thus when\n(\\theta=1) it models the length of the first sequence, which makes it equivalent\nto the geometric distribution. When (\\theta=0.5), it models the combined length\nof the first <em>two</em> sequences, etc.</p>\n<p>To make this make sense, we have to consider what happens to the hypothetical\ncoin as we increase (r) (decrease (\\theta)). We meddle with the probability of\nthe coin so that the expected total number of tails in the sequences stays the\nsame. Thus, when we increase (r), we also increase the probability, so that\nheads are more likely, and thus the length of individual sequences of tails\ndecrease, but their combined length stays the same. Everything else we are going\nto say works even if we don’t do that, but by keeping the expected total number\nof tails the same, we can see more clearly how the variance increases, as in the\nabove plots.</p>\n<p>When we get into lower (\\theta) (or higher (r)), we are now counting the length of a\nlarge number of sequences of tails. The variation in their lengths will sort of\ncancel out and converge to something more stable. Thus the variance of the\nnegative binomial distribution decreases as we decrease (\\theta) and make it\nless overdispersed.</p>\n<p>At some point we get to (\\theta \\rightarrow 0), which corresponds to (r\n\\rightarrow \\infty). At this point, the probability of heads (p \\rightarrow 1)\nand we have the Poisson distribution. But note what this means for our\ninterpretation of the Poisson distribution: the events we are counting are the\ntotal number of tails in an infinite number of sequences of tails when heads is\nalmost guaranteed to happen on every toss!</p>\n<p><em>Isn’t all this just nuts?</em>55 And what about the heavy-tailed region of the\nnegative binomial distribution? That’s when heads get increasingly <em>rare</em> but we\nonly count a partial sequence of tails to compensate.</p>\n<h1 id=\"modeling-baseball-runs-as-overdispersed-counts\"><a href=\"#modeling-baseball-runs-as-overdispersed-counts\">Modeling baseball runs as overdispersed counts</a></h1>\n<p>Getting back to the problem, we ask our computer to fit a negative binomial\ndistribution to the counts of runs each team scored. It tells us</p>\n<p>[\\mu = 4.3,\\;\\;\\;r=3.4]</p>\n<p>We know from the amazing intuition that an (r) of 3.4 corresponds to a (\\theta)\nof 0.30. It’s difficult to tell from (\\theta) alone how much overdispersion\nthere is66 We actually get a better idea from the variance which was more than\ntwice the mean. That’s why we started suspecting overdispersion in the first\nplace., but it corresponds roughly to the middle curve on the first plot of\nincreasing (\\theta). If we try to apply the amazing intuition further,\nwe may think of this as tossing a coin and counting the total number of tails in\nthe first 3.4 sequences of tails, when the coin is only slightly biased toward\ntails.77 We can tell the bias because (r \\lt \\mu) means the coin is biased\ntoward tails, while (r \\gt \\mu) means the coin is biased toward heads.</p>\n<p>I don’t know enough about baseball to translate this into physical processes,\nbut there must be some sort of decision point in the game that happens on\naverage 7.7 times per game, and 44 % of the time when it comes up, it triggers\nan event that stops the opposing team from scoring a run. When the event has\nbeen triggered 3.4 times, the game ends. Note that this doesn’t mean the 3.4th\nevent actually <em>causes</em> the game to end, it means only that the game ends on\naverage after 3.4 such events. Similarly, the lack of an event might not\ndirectly cause a run to be scored, it might just be that runs are correlated\nwith the absence of such an event.</p>\n<p>Anyway, all of this is irrelevant to forecasting on the question.</p>\n<h1 id=\"total-number-of-counts-for-7-games\"><a href=\"#total-number-of-counts-for-7-games\">Total number of counts for 7 games</a></h1>\n<p>We modeled the count for one team in one game as (\\mathrm{NB}(\\mu=4.3, r=3.4)).\nWe want the total number for all seven games in the knockout stages; that’s 14\ndraws from the distribution above.</p>\n<p>Now, the negative binomial distribution has a convenient identity:</p>\n<p>[x_i \\sim \\mathrm{NB}(r, p) \\Longrightarrow \\sum_i^n x_i \\sim \\mathrm{NB}(nr, p)]</p>\n<p>This says that <em>if</em> we have parametrised the negative binomial distribution with\n((r, p)) and we want the distribution of the sum of 14 such random values, we\ncan use the negative binomial distribution with parameters ((14r, p)).</p>\n<p>We don’t have ((r, p)), though, we have ((\\mu, r)). We could convert that to\n((r, p)) and then convert back, but there’s a more fun way: we can prove a\ncorresponding version of the identity for our parametrisation on our own! We\nwant to find (\\mu&#39;) and (r&#39;) such that</p>\n<p>[x_i \\sim \\mathrm{NB}(\\mu, r) \\Longrightarrow \\sum_i^n x_i \\sim \\mathrm{NB}(\\mu&#39;, r&#39;)]</p>\n<p>We start with the parameter transformation</p>\n<p>[p(\\mu, r) = \\frac{r}{r + \\mu}]</p>\n<p>which lets us rephrase the first identity to</p>\n<p>[x_i \\sim \\mathrm{NB}(r, p(\\mu, r)) \\Longrightarrow \\sum_i^n x_i \\sim \\mathrm{NB}(nr, p(\\mu&#39;, nr))]</p>\n<p>For this to be true still, we need to find the value of (\\mu&#39;) that makes\n(p(\\mu&#39;, nr) = p(\\mu, r)). That means we need to make it so that</p>\n<p>[\\frac{r}{r + \\mu} = \\frac{nr}{nr + \\mu&#39;}]</p>\n<p>This happens when (\\mu&#39; = n\\mu) because then we get</p>\n<p>[\\frac{nr}{nr + \\mu&#39;} = \\frac{nr}{nr + n\\mu} = \\frac{nr}{n(r + \\mu)} = \\frac{r}{r + \\mu}]</p>\n<p>Thus, the identity we are looking for is simply</p>\n<p>[x_i \\sim \\mathrm{NB}(\\mu, r) \\Longrightarrow \\sum_i^n x_i \\sim \\mathrm{NB}(n\\mu, nr)]</p>\n<p>How convenient.88 The convenience makes it even more strange that I couldn’t\nfind anyone stating this version of the identity. I must have overlooked\nsomething.</p>\n<p>This means the number of runs scored in the knockout stages of the 2026 wbc is\nprobably distributed as (\\mathrm{NB}(\\mu=14 \\times 4.3, r=14 \\times 3.4)). That\ngives us a forecast of 42–80 at 90 % confidence.</p>\n<p>The new interval of 42–80 based on more detailed modeling is better (narrower)\nthan the baseline we had at 35–87. That’s what we hoped for. Now that we\napproached it as the sum of 14 independent-but-lumpy counts, we both used more\ndata to estimate the variation, and we average out some of the noise when we sum\nit all together.</p>\n<p>Go forth and use the negative binomial distribution when appropriate!</p>\n<h1 id=\"appendix-a-on-the-independence-of-run-counts-between-teams\"><a href=\"#appendix-a--on-the-independence-of-run-counts-between-teams\">Appendix A: On the independence of run counts between teams</a></h1>\n<p>In everything above, we treated the run counts from both teams in a game as\nindependent of each other. I could hear an argument that maybe the run counts\nfor the two teams in the same game are not independent. It would make sense\nfor them to be negatively correlated: a good team scores more runs <em>and</em>\nprevents their opponent from doing the same. That would be bad for us, because\nthat would increase the variation in the count, making our forecast overconfident.</p>\n<p>Fortunately, if anything, the correlation appears to be positive. Though it is\nso weak that fitting the negative binomial model to the number of runs per\ngame rather than per team gives the same forecast interval at 90 % confidence.</p>\n<h1 id=\"sidenotes\">Sidenotes</h1>\n<p><a href=\"#fnr.1\">1</a></p>\n<p>I only got around to finishing this article now, sorry.</p>\n<p><a href=\"#fnr.2\">2</a></p>\n<p>Suggestion: use <a href=\"https://entropicthoughts.com/estimating-standard-deviation-from-timeseries-plot\" rel=\"nofollow ugc noopener\">the spc hack</a> to compute this in your\nhead.</p>\n<p><a href=\"#fnr.3\">3</a></p>\n<p>Hey, this is already a reasonable forecast for the actual\noutcome. I wonder where the community got 70 runs from.</p>\n<p><a href=\"#fnr.4\">4</a></p>\n<p>Which is a significant difference, as can be revealed\nby <a href=\"https://entropicthoughts.com/chi-square-from-fundamentals\" rel=\"nofollow ugc noopener\">a chi-square test</a>.</p>\n<p><a href=\"#fnr.5\">5</a></p>\n<p>And what about the heavy-tailed region of the\nnegative binomial distribution? That’s when heads get increasingly <em>rare</em> but we\nonly count a partial sequence of tails to compensate.</p>\n<p><a href=\"#fnr.6\">6</a></p>\n<p>We actually get a better idea from the variance which was more than\ntwice the mean. That’s why we started suspecting overdispersion in the first\nplace.</p>\n<p><a href=\"#fnr.7\">7</a></p>\n<p>We can tell the bias because (r \\lt \\mu) means the coin is biased\ntoward tails, while (r \\gt \\mu) means the coin is biased toward heads.</p>\n<p><a href=\"#fnr.8\">8</a></p>\n<p>The convenience makes it even more strange that I couldn’t\nfind anyone stating this version of the identity. I must have overlooked\nsomething.</p>","headings":[{"level":1,"text":"Baseline forecast","id":"baseline-forecast"},{"level":1,"text":"Finding underlying structure","id":"finding-underlying-structure"},{"level":1,"text":"Interlude on the negative binomial distribution","id":"interlude-on-the-negative-binomial-distribution"},{"level":2,"text":"Amazing intuition nobody cares about","id":"amazing-intuition-nobody-cares-about"},{"level":1,"text":"Modeling baseball runs as overdispersed counts","id":"modeling-baseball-runs-as-overdispersed-counts"},{"level":1,"text":"Total number of counts for 7 games","id":"total-number-of-counts-for-7-games"},{"level":1,"text":"Appendix A: On the independence of run counts between teams","id":"appendix-a-on-the-independence-of-run-counts-between-teams"},{"level":1,"text":"Sidenotes","id":"sidenotes"}]}}