连续两周的党内初选，两座摇摆州的左翼进步派候选人接连遭遇爆冷失利，这一结果给政坛观察者上了关键一课： 当下民意风向极不稳定，各类预测参考都存在巨大水分！
密歇根州与威斯康星州均缺少高质量权威民调，各大民调均值显示进步派候选人大幅领先，但这类民调大多带有党派立场倾向。预测交易市场的投注者也跟风押注，笃定看好这些进步派候选人。民调与预测市场共同营造出 “胜券在握” 的假象，直到初选开票才真相大白。
然而实际选票差距极小，胜负悬念拉满。 威斯康星州初选：热门候选人洪弗朗西斯卡以微弱劣势落败
预测平台 Kalshi 数据显示，直到威斯康星州长民主党初选开票当日（周二），州议员、民主社会主义者洪弗朗西斯卡的获胜概率仍高达 95％。（CNN 与 Kalshi 存在合作，会引用其数据报道重大选举，但 CNN 采编人员严禁参与预测市场投注。）另一平台 Polymarket 同样一边倒看好洪弗朗西斯卡。
最终，建制派候选人、密尔沃基县行政长官戴维・克劳利以毫厘之差险胜洪弗朗西斯卡。极具戏剧性的是，克劳利此前曾宣布退选，之后又重新参选。截至周三上午，两人选票差距不足 5000 张，且二人得票率均未突破 40％。另外两名知名候选人曼德拉・巴恩斯、萨拉・罗德里格斯虽在选举日前退选，但退选时间过晚，名字仍留在选票上，两人合计瓜分约 8％ 选票。
美国东部时间晚 9 点左右，选情胶着已成定局，Kalshi 平台投注者开始转向押注克劳利获胜，当晚双方胜负概率反复拉锯。CNN 直至凌晨 3 点才正式判定克劳利胜出。 密歇根州初选：热门进步派仅险胜 1 个百分点
一周前的密歇根州，另一知名左翼进步人士阿卜杜勒・赛义德此前在多份民调中领跑。这些民调均未达到 CNN 采信标准，却深深影响了预测平台投注者。初选前夜，Kalshi 给出他的获胜概率超 97％；计票全程，其胜率从未低于 94％。
最终赛义德虽然拿下初选，但得票率不足 49％，仅比温和派对手高出 1 个百分点。
两次爆冷背后，Kalshi 市场极高的获胜预期，根源都是偏向进步派的党派民调。两次选举前最新出炉的民调均由带有党派立场的机构制作。高质量中立民调缺位的情况下，舆论氛围主导了社交媒体信息茧房，营造出进步派候选人势不可挡的错觉。 威斯康星州爆冷背后的深层原因
威斯康星州唯一一份达到 CNN 采信标准的中立民调，是马凯特大学 7 月末发布的调研。这份民调中，洪弗朗西斯卡确实领先两位数，在表态会参与民主党初选的登记选民中支持率达 46％。
但关键在于，当时选情仍处于剧烈变动期：民主党建制派尚未警惕洪弗朗西斯卡，也没有统一扶持克劳利。
马凯特大学民调发布时，约三分之一选民尚未确定投票对象；当时民调第二名巴恩斯，在民调发布次日就宣布退选。候选人频繁进退，本身就会大幅增加民调精准预测的难度。
最终赢家克劳利也是近期才重新参选，而舆论针对洪弗朗西斯卡过往争议立场的集中质疑，在民调出炉后才刚刚发酵。
CNN 民调主编阿里尔・爱德华兹 － 莱维表示：“很多时候民调只能反映出选情充满不确定性、选票流向随时可能反转。” 换言之，这份高质量民调出炉时，竞选拉锯战才刚刚进入白热化阶段。
另外需要厘清概念：预测市场显示洪弗朗西斯卡 95％ 胜率，不等于她能拿到 95％ 选票，只是绝大多数投注者依据民调、往期选举趋势、候选人造势热度等信息，判断她赢面更大。
各大预测平台运营方却大肆宣传自家是 “真相机器”，刻意简化复杂现实，进一步误导大众。 平台方回应：预测失误本就是市场正常现象
Kalshi 平台创始人表示，被看好的候选人意外落败，恰恰是该市场机制的正常特性。
联合创始人塔雷克・曼苏尔在 X 平台发文：“在‘预测市场翻车’的新闻满天飞之前，要明白 5％ 的获胜概率不代表这件事不会发生，而是平均每 20 次就会出现一次。如果 5％ 概率的事件永远不会成真，这套市场体系才是彻底失效。”
Kalshi 称这次事件是大众的一堂概率认知课。 平台发言人杰克・苏奇周三向 CNN 表示：“当下社会越来越缺少辩证思考，而 Kalshi 最重要的价值，就是教会大众用概率思维看待事物。”Polymarket 平台则拒绝发表评论。 民调质量才是判断选情的核心
密歇根、威斯康星两州近年多次出现民调严重失准的案例，业内民调机构公认这两个州极难精准调研。两州选民投票立场摇摆性极强，且两州均不实行党派选民登记制度 —— 现代民调依赖选民档案数据建模，无党派登记会直接拉高调研误差。
这也意味着，想要准确判断选举走向，足量、高质量中立民调不可或缺。如今愿意投入资源做权威民调的媒体越来越少，CNN 不会仅凭单一民调下结论，在条件允许时会整合所有达标民调算出综合均值，以此评估总统支持率、大型选情走势。 展望 11 月中期选举：局势变数极大
看待 11 月大选民调，要读懂数据传递的信息，更要看懂数据无法反映的盲区。 SSRS 机构为 CNN 制作的最新民调显示，中期选举选民整体对民主党偏好度领先共和党 8 个百分点。但这份参考价值有限：选民不会笼统地为两党投票，而是针对具体选区、特定候选人做出选择。美国多个州还通过选区划分操作（杰利蝾螈）偏袒某一政党，将立场趋同选民集中划分至单一选区。
过去两周两场初选给我们明确启示：距离 11 月大选还有漫长周期，民意风向随时可能逆转。后续会涌现海量立场分化的民调，其中不少由带有党派诉求的机构制作，参考价值有限。
整个夏季，Kalshi 预测市场中民主党拿下众议院控制权的概率持续走高，当前已超 85％。看似优势巨大，但依然给共和党留有翻盘空间。按平台概率测算，共和党保住众议院多数席位的概率，甚至高于克劳利当初赢下威斯康星州长民主党初选的概率。我们要理性读懂数字含义：高额投注代表大众认为民主党胜算更高，而非胜局已定；参议院控制权的两党获胜概率则基本持平。
预测市场是有趣的参考工具，但它的判断上限，永远等同于当下主流舆论的认知水平，不能当作绝对定论。
What a safe－bet candidate’s loss means for Democrats’ odds in November
Two successive weeks of primaries brought two successive shocks for progressive candidates in purple states and carry an important lesson for people watching politics：
It’s noisy out there！
With little high－quality polling in either Michigan or Wisconsin， polling averages showing big leads for progressive candidates were driven largely by partisan polling． And prediction market bettors followed suit， going all－in for those candidates． The two combined to create the mirage of a sure thing ahead of primary day．
But actual voters served up tight margins．
Francesca Hong， the democratic socialist and state assembly member in the Democratic primary for Wisconsin governor， had a 95％ chance of victory up until Tuesday， primary day， according to theprediction site Kalshi． （CNN has a partnership with Kalshi and uses its data to cover major events． But CNN editorial employees are not allowed to use prediction markets．） There are other prediction platforms， includingPolymarket， that similarly favored Hong．
Hong ended up losing the race by thetightest of marginsto David Crowley， the establishment candidate and Milwaukee County executive who， incredibly，dropped out of the race before rejoining it． As of Wednesday morning， fewer than 5，000 votes separate the two． And neither reached 40％ in the race． Two other prominent candidates — Mandela Barnes and Sara Rodriguez — dropped out ahead of Election Day， but not soon enough to remove their names from the ballot， and together won about 8％ of the total votes cast．
By around 9 p．m． ET， it was clear the vote was going to be close， and traders on Kalshi were predicting a Crowley victory． The odds see－sawed through the night． CNN did not project a Crowley victory until the 3 a．m． ET hour．
In Michigan a week earlier， another proud progressive， Abdul El－Sayed， seemed like a lock with a lead in recent polls， none of which met CNN’s standards for reporting， but which may have affected bettors on prediction sites． It all led up to a more than97％ chance of victoryon primary eve， according to the market on Kalshi． His odds never dropped below 94％ as votes were being counted．
El－Sayed ended up winning， but with less than 49％ of the vote， just one percentage point ahead of his moderate rival．
In both cases， the near－certainty of bettors on Kalshi was likely fueledby polls that suggested leads for Hong and El－Sayed． But in both cases， the most recent polling had been conducted largely by partisan outfits． In the absence of more high－quality polling， vibes took over and fed the social media echo chamber and a sense of momentum for progressive candidates in particular．
What happened in Wisconsin？
In Wisconsin， the most recent independent poll that met CNN’s standards had been conducted in late July byMarquette University． In that poll， Hong was indeed ahead by double digits， holding 46％ support among registered voters who said they would vote in the state’s Democratic primary．
Importantly， the race was still in flux． The Democratic establishment had not yet rung alarm bells about Hong and coalesced around Crowley．
About a third of voters were uncertain about their choice when that Marquette poll was conducted． Barnes held second place in the poll， and dropped out just a day after the poll was released． It’s also hard to accurately poll a race where the candidates keep changing．
The ultimate winner， Crowley， had only recently reentered the race． And public scrutiny of some of Hong’s more controversial previous positions was just getting underway．
“Sometimes what polling tells you is that there’s a lot of uncertainty and room for shifts in a race，” said CNN’s polling editor， Ariel Edwards－Levy．
In other words， the campaign was still very much underway when the last high quality polling was done．
The fact that Hong had a 95％ chance of winning on prediction markets never meant that she was going to get 95％ of the vote or anywhere near it． It meant that most bettors thought she would win based on what they were seeing — polling， trends in other recent elections， their reads on the mojo of the candidates and whatever else fed the wisdom of the crowd．
In addition to those limits， it might not help when Polymarket and Kalshi executives tout their products as “truth machines，” when the reality is far more nuanced．
The crowd was wrong
Kalshi’s founders say favored candidates occasionally losing is actually a feature of the platform．
“Before the ‘prediction markets got it wrong’ headlines roll in： a 5％ probability doesn’t mean it won’t happen． It means it should happen 1 in 20 times． If 5％ candidates never won， the markets would be broken，” Kalshi co－founderTarek Mansourwrote on X．
In fact， Kalshi argues this is a teachable moment．
“Teaching people to think probabilistically in a world that is quickly losing nuance is one of Kalshi’s most important impacts，” Kalshi spokesman Jack Such told CNN on Wednesday． Polymarket declined to comment．
Polling quality matters
Michigan and Wisconsin are home to some of the highest－profile recent election polling misses， and pollsters agree they are notably places that are hard to poll well． Voters in both states have proved reliably swingy， and neither state has partisan voter registration， which makes difficult the type of modeling on voter file data that improves modern polling．
It all means that the quantity of quality polling is important in knowing how confident to be about the state of any race， particularly as fewer news organizations conduct polls． CNN does not rely on any single poll， but does compute apoll of polls— essentially an average of polls that meet its standards — where possible to assess presidential approval ratings and， as elections approach， for major contests．
Looking ahead to November， polling should be viewed for what it does say rather than what it does not．A recent CNN Pollconducted by SSRS showed that voters preferred Democrats by an 8－point margin heading into the November midterm election． That number is only so helpful， however， since people aren’t voting generally for Republicans or Democrats； they’re voting in a specific race for a specific candidate． House seats in many states have also been gerrymandered to benefit one party or the other， stuffing like－minded voters into certain districts．
What we’ve learned over the past two weeks is that November is still an eternity away and momentum can shift．
There are going to be a lot of polls out there that may point in different directions， a lot of them conducted by groups with a partisan agenda．
On the Kalshi prediction market， the odds of Democrats taking control of the House has been growing all summer and currently sits at more than 85％， which is a high degree of confidence， but also leaves a lot of room for Republicans． They actually have a better chance of holding the House majority than Crowley had of winning the Democratic primary for governor in Wisconsin， under Kalshi’s odds． It’s important to remember what these numbers mean： People who are placing monetary bets think the safer bet is on Democrats winning． In the Senate， the odds of one party or the other winning control are about even．
Those markets are an interesting tool， but they’re only as good as the conventional wisdom．
