纳特·西尔弗:特朗普有势头吗

文摘   2024-10-22 14:30   北京  

10月18日,有一些民调首度认为特朗普的支持率反超哈里斯。作为最了解美国大选民调的Nate Silver,他却认为,应该审慎看待这一情况。他写过几本有关民调的书,包括《信号与噪音:为何大多数预测会失败但另一些却不会》、《边缘地带:成功的赌徒和风险承担者是如何思考的》等。现将此文翻译如下(由知名AI软件DeepL协助翻译):
Well, it happened.
好吧,这发生了。
It was almost inevitably going to happen at some point, given how close the forecast has been to 50/50 since Kamala Harris took over for Joe Biden. But for the fourth time this cycle, the streams have crossed, and the nominal favorite in the race has changed. Thanks in part to one of the first high-quality national polls in weeks to show him ahead — yes, it’s from Fox News, but Fox News polling has no history of GOP bias — Donald Trump now has a 50.2 percent chance of winning the Electoral College.
自从卡玛拉·哈里斯接替乔·拜登以来,预测结果一直接近50/50,因此这个情况的发生几乎是不可避免的。 但这已经是本轮选举的第四次交锋,竞选中的名义热门人选也发生了变化,这或要归功于高水平的福克斯民调。 唐纳德·特朗普现在有50.2%的机会赢得选举团。
That’s not much different from Trump’s 49.4 percent chances yesterday: the difference is that he wins the Electoral College an additional 1 time out of 125. But Trump is past the 50-yard line — and above Harris’s chances, which are 49.5 percent (Note 1-There’s also a 0.3 percent chance of a 269-269 Electoral College tie) — for the first time since Sept. 19. (Harris remains a clear but not overwhelming favorite in the popular vote, but that’s not how American elections are decided.)
这与特朗普昨天49.4%的胜选几率并无太大区别:区别在于他在125次选举中多赢得1次选举团投票。 但特朗普自9月19日以来首次超过了50码线——超过了哈里斯的机会,即49.5%。(注1:还有0.3的可能性,选举团出现 269-269 票平局。) (哈里斯在普选中的优势依然明显,但并非压倒性优势,但这不是美国选举的决定方式)。
To be open about my editorial thought process, it can be hard to know how much to emphasize these small shifts. Outside of major events like debates, the model typically does not move very much from day to day. And there’s no special significance to the 50-yard line. A change from 49.4 percent to 50.2 percent does not particularly matter more than one from 43.1 percent to 43.9 percent or 52.5 percent to 53.3 percent.
坦率地说,在我的编辑思维过程中,很难知道该在多大程度上强调这些微小的变化。 除辩论等重大事件外,模型通常每天都不会有太大变化。 50 码线也没有什么特别的意义。 从 49.4%到50.2%的变化并不比从43.1% 到43.9%或从52.5%到53.3%的变化更重要。
I’m also aware that most coverage on the polling beat tends to overemphasize these small changes. People treat a 51/49 forecast very differently from a 49/51 forecast when they shouldn’t. There are also some partisan asymmetries here: Republicans tend to gloat about polling leads, while Democrats tend to live in fear they’ll evaporate. And Republicans more actively cultivate narratives that their candidate has momentum in the race.
我也意识到,大多数关于民调节奏的报道往往过分强调这些微小的变化。 人们对待 51/49 和 49/51 预测的态度大相径庭,而这是不应该的。 这里还存在一些党派的不对称: 共和党人倾向于为民调领先而沾沾自喜,而民主党人则往往生活在对领先优势消失的恐惧之中。共和党人会更积极地制造自己的候选人在竞选中势头强劲的论调。
But life is full of arbitrary numerical thresholds: hitting 50 home runs is a much bigger deal than hitting 49, and your 40th birthday is a much bigger deal than your 39.
但是,生活中充满了武断的数字门槛:击出50支本垒打比击出49支要重要得多;您的40岁生日比39岁的要重要得多。
And it’s not like these small changes have offset one another: lately, the daily shifts have mostly worked in Trump’s favor. He bottomed out at a 41.7 percent win probability on Sept. 27, about two weeks after his debate with Kamala Harris. Since then, his odds have improved day-over-day 12 times, decreased 7 times, and been unchanged once. So that’s not an overwhelmingly consistent pattern, but it starts to add up. I want to avoid the sort of boiling frog syndrome where you ignore a trend because it happens slowly and incrementally.
这些微小的变化并没有相互抵消:最近,每天的变化大多对特朗普有利。 9月27日,即他与卡玛拉·哈里斯辩论两周后,他的胜选概率跌至41.7%的谷底。 从那时起,他的胜选几率12次逐日上升,7次下降,1次不变。 因此,这并不是一个压倒性的一致模式,但它开始累加起来。 我希望避免“温水煮青蛙综合症”,即因为趋势发生得缓慢且渐进而忽略它。
Plus, for what it’s worth, the Silver Bulletin model has been slightly more explicit about showing a trend toward Trump over the past week or so than other models — matching prediction markets in that respect. Although, prediction markets are probably going too far in this case and can overstate the importance of momentum. (Note 2-They can also be influenced by the presence of whales, i.e. well-bankrolled individual bettors who can buy up lots of contracts. )
另外,值得注意的是,在过去一周左右的时间里,“西尔弗简报”模型比其他模型更明确地显示了特朗普的趋势,在这方面与预测市场不谋而合。 不过,在这种情况下,预测市场可能走得太远,会夸大势头的重要性。(注2:它们还可能受到“鲸鱼”的影响,即资金雄厚的个人投注者可以买入大量合约。
I’ve already used the “m-word” twice — momentum — without formally defining it, and the formal definition is important. Does the fact that Trump’s chances have increased from 42 percent to 50 percent imply that they’ll continue to increase? Hold that thought, though, because I think I owe you a rundown of where Trump’s odds have improved. There are some pretty consistent patterns at the state level:
我已经使用了“m-单词”两次——“势头(momentum)——但尚未对其进行正式定义,而正式定义是很重要的。 特朗普的胜算从42%上升到 50%,这是否意味着他的胜算会继续上升? 先别这么想,因为我想我还欠你一份特朗普胜算提高的详细说明。 在州一级有一些相当一致的模式:
It’s basically the Blue Wall states. Trump has gone from being almost a 2:1 underdog in Michigan, Wisconsin and Pennsylvania at his nadir to close to 50/50 now. Outside of the Blue Wall (and Florida) the changes are incremental. But the Blue Wall states have a combined 61 percent chance of being the tipping-point state, so they really take a bite out of Harris’s Electoral College chances. The closest thing to a bright spot for Harris is North Carolina, which has surpassed Wisconsin in tipping-point odds and could give Harris an out if she loses one of the Blue Wall states (Note 3-If North Carolina substitutes for Pennsylvania, she’d also have to win Nevada — but it would substitute for Michigan or Wisconsin on its own.) (but probably not more than one of them).
基本上就是“蓝墙州”。 特朗普在密歇根州、威斯康星州和宾夕法尼亚州的劣势从最低谷时的几乎2:1变为现在的接近50/50。 “蓝墙州”(和佛罗里达州)之外的变化是渐进的。 但 “蓝墙州”有61%的几率成为关键州,因此它们确实削弱了哈里斯进入选举团的机会。 哈里斯最接近亮点的州是北卡罗莱纳州,该州的触发机会率已超过威斯康辛州,如果哈里斯输掉“蓝墙州”中的一个州(注3:如果北卡罗来纳州取代宾夕法尼亚州,她还必须赢得内华达州——但它本身就可以取代密歇根州或威斯康星州。(但可能不会超过一个),她可能会有一线生机。
So what is “momentum” anyway?
那么,“势头”到底是什么?
I wrote about momentum in polling way back in 2010. Here’s how I described it then:
早在 2010 年,我就写过关于民意调查势头的文章。 我当时是这样描述的:
Turn on the news or read through much of the analysis put out by some of our friends, and you’re likely to hear a lot of talk about “momentum”: the term is used about 60 times per day by major media outlets in conjunction with articles about polling.
打开新闻或阅读我们一些朋友的分析文章,你很可能会听到很多关于“势头”的讨论:主要媒体每天在发表有关民调的文章时使用这个词约 60 次。
When people say a particular candidate has momentum, what they are implying is that present trends are likely to perpetuate themselves into the future. Say, for instance, that a candidate trailed by 10 points in a poll three weeks ago — and now a new poll comes out showing the candidate down by just 5 points. It will frequently be said that this candidate “has the momentum”, “is gaining ground,” “is closing his deficit,” or something similar.
当人们说某位候选人势头强劲时,他们的意思是目前的趋势很可能会延续到未来。 比如说,某候选人三周前在民调中落后10个百分点,而现在新的民调显示该候选人只落后5个百分点。人们经常会说,这位候选人“势头强劲”、“正在取得优势”、“正在缩小差距”或类似的说法。
Each of these phrases are in the present tense. They create the impression that — if the candidate has gone from being 10 points down to 5 points down, then by next week, he’ll have closed his deficit further: perhaps he’ll even be ahead!
这些短语都是现在时。它们给人的印象是——如果候选人从落后10个百分点到落后5个百分点,那么到下周,他就会进一步缩小差距:也许他甚至会领先!
Emphasis added. The point is that most people tend to bet (literally in the case of prediction market traders) on trends continuing. If Trump has gone from 42 percent to 50 percent, in other words — already bad news for Harris — “momentum” implies that he’ll keep moving up: perhaps by Election Day, he’ll be at 60 percent, for instance. And that would be really bad news for Harris, the point where we could no longer honestly describe the race as a toss-up.
需要强调一下。 问题的关键在于,大多数人倾向于押注(在预测市场交易商的情况下是字面意思)趋势的持续。 换句话说,如果特朗普的支持率从 42% 上升到 50%(这对哈里斯来说已经是个坏消息),那么“势头”就意味着他会继续上升:例如,到投票日时,他的支持率可能会达到 60%。 而这对哈里斯来说将是一个真正的坏消息,到那时我们就不能再诚实地将这场竞选描述为势均力敌了。
Except that, as the 2010 article concluded, polling generally does not show this pattern, at least not over the medium term. Instead, as an empirical matter, polling averages more closely resemble random walks.
只不过,正如2010年的文章所总结的那样,民调一般不会显示出这种模式,至少在中期内不会。 相反,根据经验,民调平均值更接近随机漫步。
This is the same way, incidentally, that the stock market behaves, at least most of the time. If the price of Nvidia rose from $135 to $140 today, and that predictably implies it will rise to $145 tomorrow, then of course you should buy it. Of course, there are “momentum traders” that bet on exactly this hypothesis — but usually, it doesn’t produce any excess return. The market is too efficient. If everyone knew Nvidia would rise to $145 tomorrow, that would already be “priced in” to today’s price. That’s what the theory says, anyway.
顺便说一句,这与股票市场的行为方式是一样的,至少在大多数时候是这样。 如果今天英伟达的价格从135美元涨到了140美元,这就意味着明天会涨到145美元,那么你当然应该买入。当然,也有一些“动量交易者”正是在这种假设上下注——但通常不会产生任何超额收益。市场太有效率了。如果每个人都知道英伟达明天会涨到145美元,那么今天的价格就已经是“定价”了。 反正理论上是这么说的。
The reality may be slightly more complicated, in part because price movements may be asymmetric. If there’s an 80 percent chance that NVDA rises to $145 tomorrow but a 20 percent chance it crashes to $120, the expected value of the stock is $140, but you’ll still win your bet and look smart 80 percent of the time.
实际情况可能稍微复杂一些,部分原因是价格走势可能不对称。如果英伟达明天上涨到145美元的概率为 80%,但下跌到120美元的概率为20%,那么该股票的预期价值就是140美元,但你仍然会在80%的时间里赢下赌注,而且看起来很聪明。
And maybe you’ll decide you can time the market or get out just at the right moment, or you’re a well-reputed trader who can influence other traders. We’ve seen more and more of this herd behavior in recent years in both traditional and nontraditional (e.g. crypto) assets, and some traders at political prediction markets do the same thing. They’ll buy Trump shares not because they necessarily expect Trump’s fundamental value to improve (for instance, buy getting a bunch of strong polls) but because they think they can read trader sentiment or even influence it — helping to develop a narrative that Trump has momentum.
也许你会认为自己能把握市场时机,或者能在正确的时刻出局,或者你是一位声誉卓著的交易者,能影响其他交易者。 近年来,我们在传统和非传统(如加密货币)资产中看到了越来越多的这种羊群行为,政治预测市场上的一些交易员也做着同样的事情。 他们会买入特朗普的股票,并不是因为他们一定会预期特朗普的基本价值会提高(例如,买入获得大量强劲民调的股票),而是因为他们认为自己可以读懂交易员的情绪,甚至可以影响交易员的情绪——帮助形成特朗普势头强劲的说法。
A good election forecast avoids momentum
一个好的选举预测是避免造势的
The Silver Bulletin election forecasts are explicitly designed to be momentumless. In other words, the best prediction of what the forecast says tomorrow is what it says today; the technical term for this is that it avoids autocorrelation.
“西尔弗简报”的选举预测是明确设计为不具势头的。换句话说,预测明天的最佳预测结果就是今天的预测结果;这里的专业术语是避免自相关性。
This is easier said than done, however. For one thing, it involves carefully tuning how sensitive the polling averages and model are to new polls. If they’re too sensitive, they’ll bounce around too much: Trump going from +1 in Arizona last week to +2 this week implies that he’ll revert back to +1.5 next week. This is mean reversion or negative autocorrelation. On the other hand, models can be too sluggish to react to new polls. If Trump rising from +1 to +2 implies that the polling average next week will be +3, this is positive autocorrelation and you’re leaving predictable information (that Trump will continue to rise) on the table.
然而,说起来容易做起来难。 首先,这需要仔细调整民调平均值和模型对新民调的敏感程度。 如果过于敏感,它们就会大幅波动:特朗普上周在亚利桑那州的支持率为+1,而本周为+2,这意味着他下周将回到+1.5。 这就是均值回归或负自相关。 另一方面,模型对新民调的反应可能过于迟钝。 如果特朗普的民调从+1升至+2,意味着下周的民调平均值将为+3,这就是正自相关,你就把可预测的信息(特朗普将继续上升)留在了台面上。
For what it’s worth, the Silver Bulletin polling averages tend to be more sensitive than most, because one of the most important factors I sought to optimize for when tuning their parameters was minimizing autocorrelation — that is, having neither positive autocorrelation nor negative autocorrelation. While laypeople tend to overemphasize a forward trend — their “mental models” are too sensitive to new information, leading them to gyrate back and forth like a fan at a tennis match — people who build and publish models professionally probably err toward the opposite mistake. They think it makes them look silly if their forecast is bouncing around a lot, so they build models that “ease into” changes more than is optimal if you want to maximize predictive accuracy.
值得注意的是,“西尔弗简报”的民调平均值往往比大多数其他民调更敏感,因为在调整其参数时,我试图优化的最重要因素之一就是尽量减少自相关性,即既没有正自相关性,也没有负自相关性。 普通人往往会过分强调前进趋势——他们的“心智模型”对新信息过于敏感,导致它们像网球比赛中的球迷一样来回摆动——而建立和发布专业模型的人可能会犯相反的错误。 他们认为,如果自己的预测经常反反复复,会让自己看起来很傻,所以他们建立的模型“容易接受”变化,而不是想要最大限度地提高预测准确性的最佳模型。
However, there are several complications. One is that you have to decide what time period you care about. Do I want a polling average that best predicts tomorrow’s polling average — or, say, best predicts the polling average 30 days from now?
不过,这也有几个复杂的问题。 其一,您必须决定您关心的时间段。 我是想要一个能最好地预测明天民调平均值的民调平均值,还是想要一个能最好地预测 30 天后民调平均值的民调平均值?
These won’t necessarily give you the same answer. Let’s say there’s an event like a debate. Kamala Harris wins the debate. There are a couple of new national polls, and she rises from a 2-point lead to a 3-point lead in national polls. It might be that the best prediction for her polling average next week is a 4-point lead, because you expect her to continue to get favorable media coverage, for more people to watch clips of her good debate, and so on.
这些不一定会给您相同的答案。 我们假设有一场辩论。卡玛拉·哈里斯赢得了辩论。 有几项新的全国民调显示,她在全国民调中的领先优势从2个百分点上升到3个百分点。 对她下周民调平均值的最佳预测可能是领先4个百分点,因为你预计她将继续得到媒体的有利报道,有更多人观看她精彩辩论的片段,等等。
But it might also be that the best prediction for her polling average in 30 days is for it to revert back to a 2-point lead, because bounces from events like debates and conventions can be short-lived. The polling average will need to be very aggressive if you have a short-term focus but quite conservative if you’re looking toward the medium-to-long term. At Silver Bulletin, we seek a middle ground; our polling averages are designed to look two weeks out — that is, the polling average today should optimally predict what new polls will say two weeks from now, except at the end of the race when there’s no time left for reversion to the mean, and we use a more aggressive setting. It’s possible we’ll understate momentum over the very short term, however.
但是,对她的民调平均值在30天内的最佳预测可能是恢复到2个百分点的领先优势,因为辩论和大会等活动带来的反弹可能是短暂的。 如果您着眼于短期,那么民调平均值就需要非常激进,但如果您着眼于中长期,那么民调平均值就需要相当保守。 在“西尔弗简报”,我们寻求的是中庸之道;我们的民调平均值旨在预测两周之后的情況,即今天的民调平均值应能最好地预测两周后的新民调結果,除非在竞选结束时没有时间还原到平均值,我们才会使用更进取的设定。不过,我们有可能在短期内低估了势头。
It’s also plausible that there are times when polls reflect and even contribute to various sorts of positive feedback loops. A candidate gets some strong polls, and that produces favorable media coverage or a bandwagon effect where voters jump to support candidates who appear to be bound for victory. For whatever sociological reason, Republicans really tend to believe in this theory, which is why they can develop a real pack mentality whenever Trump gets a strong batch of polls or the news cycle seems to be favoring him.
民调反映甚至促成各种正反馈循环的情况也时有发生。某位候选人在民调中表现突出,媒体就会对其进行有利的报道,或者会产生一种浪潮效应,选民就会跃跃欲试,支持那些似乎必将获胜的候选人。 无论出于何种社会学原因,共和党人确实倾向于相信这一理论,这就是为什么每当特朗普获得一批强有力的民调结果或新闻周期似乎对他有利时,他们就会产生一种真正的群体心态。
Indeed, these effects can be quite important in multi-candidate races like presidential primaries where voters behave tactically. In 2020, when Elizabeth Warren began slumping in polls, a lot of her voters switched to Bernie Sanders because they wanted a more progressive alternative to Joe Biden. It became impossible for Warren to dig herself out of the hole.
事实上,在总统初选等多候选人竞选中,这些影响可能相当重要,因为选民会采取策略性行为。2020 年,当伊丽莎白·沃伦的民调开始下滑时,她的很多选民转而支持伯尼·桑德斯,因为他们希望在乔·拜登之外有一个更进步派的选择。 沃伦不可能再从困境中走出。
However, my experience both as a model-builder and as someone who’s been through a lot of political news cycles is that these momentum effects tend to be overrated in two-way races like presidential general elections. That’s because, to the extent it matters at all, media sentiment shifts rapidly. “Trump rising” becomes a boring storyline after about two weeks, expectations have risen, and the media will begin to look for excuses to declare a “Kamala comeback”. There are still 18 days left to go in the campaign, and if Harris wins, it’s possible that we’ll look back on this period and say that Trump peaked too soon — that the last and most important turn of momentum favored Harris, not Trump.
然而,作为一个模型建立者和一个经历过许多政治新闻周期的人,我的经验是,在总统大选这样的双向竞选中,这些势头效应往往被高估。 这是因为,就其重要程度而言,媒体情绪会迅速转变。大约两周后,“特朗普崛起”就会成为一个无聊的故事情节,人们的期望值也会上升,媒体就会开始寻找借口宣布“卡玛拉卷土重来”。竞选活动还剩下 18 天,如果哈里斯获胜,我们在回顾这段时期时有可能会说,特朗普过早地达到了顶峰——最后也是最重要的势头转折有利于哈里斯,而不是特朗普。
All of this is further complicated by partisan non-response bias and pollster herding. The former phenomenon is that voters may be more likely to respond to polls when their candidate is getting favorable news coverage. The latter issue is that mediocre pollsters may decide it’s safe to publish more favorable numbers for a candidate when he’s in the midst of a favorable news cycle or when other pollsters are showing him doing well. If Trump is perceived to be on an upswing, they’ll publish that Trump +4 in Pennsylvania, but if Harris is perceived to have momentum instead, they’ll tinker with their turnout model until it shows a tie instead. However, this is artificial momentum. What matters in the end is votes, not polls.
党派非回应偏差和民调机构羊群效应使所有这些问题变得更加复杂。 前一种现象是,当候选人得到对其有利的新闻报道时,选民可能更愿意对民调做出回应。 后一个问题是,平庸的民调机构可能会认为,当候选人处于有利的新闻周期中或其他民调机构显示其表现良好时,公布对其更有利的数据是安全的。如果特朗普被认为处于上升阶段,他们就会公布特朗普在宾夕法尼亚州的支持率为+4,但如果哈里斯被认为势头强劲,他们就会修改投票率模型,直到显示平局为止。 然而,这是人为的势头。最终重要的是选票,而不是民调。
Finally, our models have a notion of the “fundamentals” based on factors like the economy and incumbency. For instance, our current economic index implies that Harris “should” win the popular vote by roughly 1 percentage point. Although the fundamentals do not currently get very much weight in our forecast — they are phased out over the course of the campaign — this is bad news for Harris as it implies that her numbers were previously inflated when she led Trump by as much as 3 or 3.5 points in national polls. Looked at in this way, Trump’s recent gains in the polls are actually not “momentum” but mean reversion, the race returning to its natural baseline after Harris’s strong debate.
最后,我们的模型有一个基于经济和现任等因素的“基本面”概念。 例如,我们目前的经济指数意味着哈里斯“应该”以大约1个百分点的优势赢得普选。虽然基本面因素目前在我们的预测中并不占很大权重——在竞选过程中会逐步被剔除——但这对哈里斯来说是个坏消息,因为这意味着她之前在全国民调中领先特朗普多达3或3.5个百分点时,她的数据被夸大了。 从这个角度看,特朗普最近在民调中的优势实际上并非“势头”,而是均值回归,即在哈里斯的精彩辩论之后,竞选回到了自然基线。
I realize now that I pulled a switcheroo on you about midway through this newsletter and turned a relatively nontechnical post into an extremely wonky one — but hey, that’s what Model Talk columns are all about. The most important takeaways is are that 1) Trump has made real gains — at this point, we’re probably beyond the point where this can just be statistical noise but 2) you should be wary when you hear media talk about “momentum”: that he’s gained ground over the past two weeks does not necessarily imply that he’ll continue to do so.
我现在意识到,我在这篇时事通讯的中途耍了个花招,把一篇相对非技术性的文章变成了一篇极其古怪的文章——不过,这就是“模式之谈”专栏的意义所在。 最重要的启示是:1)特朗普取得了真正的进展——在这一点上,我们可能已经超越了这可能只是统计噪音的程度;2)当你听到媒体谈论“势头”时,你应该警惕:他在过去两周取得的进展并不一定意味着他会继续这样。
————————
简评:作者的那两句——“共和党人倾向于为民调领先而沾沾自喜,而民主党人则往往生活在对领先优势消失的恐惧之中。共和党人会更积极地制造自己的候选人在竞选中势头强劲的论调。”——跟我之前对拜登辩论的感受很相似,不是拜登多么弱,而是两党对于拜登表现的反应很有趣。
换句话说,辩论并不重要,重要的是辩论之后的造势如何。我认为拜登辩论得不错——这没用,因为这之后的团体及舆论造势他不行了。另一场辩论,大家都认为哈里斯赢了——那也没用,因为算法变了,造势未能随着算法改变而调整,导致那场辩论之后的民调显示,辩论对民调没有多大影响。
Nate Silver根据大数据指出,不仅辩论在一定程度上不重要,就连某人支持率第一次过了半数也在一定程度上不重要。就像股票或涨或跌,您就一定认为您的某支股票涨了两天就继续涨第三天吗?可不,星座运势都不能保证您高兴了两天,第三天就不会遇上水逆呢?
对我而言,他说就好比有人要过40岁生日了,会比我的39岁生日重要吗?并不,但确信无疑,因为我的敏感,以及中国传统,我一定会赋予我的40岁生日更多的意义。

熊法通则
高楼高,高楼美,我在楼下往上瞧。哎呀呀,掉了我的小花帽。
 最新文章