Sift 周预测第 44 周：14 事件 9 注净赚 2 美元，与 Eli Lilly「一份预测、两条相反交易」

信源说明：本文基于 All About AI 频道主 Filipe 一期视频整理（视频 ID SaBtmnZGBxo，全长约 13.9 分钟，英文转写 0s–836s 已完整读取，文中数字均依转写）。本频道为个人创作者+创业者频道，全部数字为作者屏幕展示的自家模型输出、持仓与盈亏，本文未做独立核实；按系列惯例以「作者称/模型展示」标注归属，ASR 疑似错词以「转写疑作」标注，文末附校正表与完整转写原文。同系列前文：《AI 预测股市第一周战绩》、《Sift 过去情景回测》、《拿自家预测产品押注中期选举》。

先说结论

频道主把自己的 AI 超预测产品从「回测与选战留档」推进到新的产品形态：每周真金白银的短线预测系列。第 44 周（10 月 26 日周一至 30 日周五）他在 Polymarket/Kalshi 上选了 14 个事件——美国 Q3 GDP、英伟达周线、黄金/布伦特/十年期收益率、ECB 与日本央行利率决议、霍尔木兹海峡航运、亚马逊与礼来财报、欧元区与德国 GDP、SPY 触及 777 等——每注 10 美元。结果是克制的：14 个里只成交 9 注（其余没等到想要的价位），PNL 约 +2 美元，作者自己的总结是「没亏，但也没赚」。本周真正的增量在结尾的礼来（Eli Lilly）案例：Sift 只出一份概率报告，由 Claude Code 运行 Opus 5.5 的 agent 把它翻译成两条方向相反、但同源的交易，并双双命中——「Sift 给概率，agent 找交易」这句话，比 +2 美元重要得多。

周预测怎么跑：只打「有分歧」的盘

作者的筛选逻辑本身就是一个可迁移的方法（120–360s）：ECB 与 BOJ（转写一处误作 BOG）的十月利率决议，模型给到 88%、95%，「基本死了」——即使与市场差 10 个点也直接跳过；亚马逊财报模型与市场「完全一致」，不下注。真正动手的是有分歧的盘：美国 Q3 GDP 增速超 3%，市场 63、模型 54，差 8 个点——他特意说明这里模型方向其实是 Yes，但他选择押 No「博更大的上行空间」（这一注后来输了，是本周反面教材）；英伟达周线收于某个价位之上，模型 66% 对市场 47%，但盘口价差 33 美分、「位置很难受」，他选择观望等价格；黄金 10 月 30 日价位，模型约 51，等回调以 52 进了 8 美元。也就是说：事件池按日历抓，下注按「分歧幅度+盘口流动性」过滤——这延续了他前几期「分歧点即卖点」的思路。

结果盘点：最佳一注与最痛一注

周六（10 月 31 日）复盘（360–600s）：最佳是「欧元区 Q3 GDP 落在 0.8–1.1 区间」的 No，35 美分买入、按 100 结算；德国 GDP 的 No「翻倍」；礼来 EPS 7.49 美元「是否超预期」的 No 中标。输掉的有：美国 GDP（方向本就与模型相反）、十年期收益率站上 4.97、黄金站上 4000；SPY 触及 777 的长射「有点不走运」——作者称以 6–7 美分左右进场（转写此段含混，疑为长射买方），SPY 全周最高约 771、随后一路滑到 756，差 6 个点没摸到线；布伦特站上 112 的一注当时尚未结算。合计 PNL 约 +2 美元。作者的态度与系列前三期一致：第一周、样本为零、留档为以后积累 track record，「下周带新数字回来」。

礼来案例：一个预测，两条相反的交易

本期信息密度最高的部分（600–836s）是作者回放的一段礼来财报交易拆解。Sift（转写通篇误作 SICT/Six Intelligence）在财报前两天给出预测：「礼来 EPS 是否超过 7.49 美元」，市场 44%、Sift 33%。按播放的拆解：第一条线索在「线」本身——礼来已宣布每股 4.57 美元的一次性费用，它就埋在 7.49 这条线里，真实的门槛约 12 美元，高到不现实→交易一：Polymarket 押 No。第二条线索在华尔街——分析师最低估计只有 5.38 美元，因为很多人没把那笔费用算足；而礼来此前连续两个季度 underlying 口径超预期约 20%→即便报表数字压线之下，实际业务看起来仍像大超预期→交易二：在 Interactive Brokers 买多礼来股票。把这两条线索组装成交易的不是模型本身，而是 agent：Claude Code 运行 Opus 5.5，读 Sift 报告后只问两个问题——「能过线吗？不能。」「会超出华尔街已被期权定价的预期吗？会。」——两个回答，两条腿。10 月 29 日礼来报出 7.48 美元，恰好低线一美分：No 一注赔率 +75%，同日股票涨约 8%（销售远超预期并上调指引）。「一份预测、两条看起来相反的交易，都对。」

机制与启发

其一，短线预测市场的 edge 不在「预测得准」，而在「分歧可解释、可下注、可复盘」。 作者的过滤器——跳过模型与市场一致的盘、只在 8 个点以上分歧处出手、进不到想要的价格就放弃（14 选 9）——把「预测」变成了一个带纪律的下注流程。启发：一周 9 注、±2 美元的样本在统计上约等于零信息，这个系列的价值只能靠几十周的公开结算记录兑现；观察指标是分歧盘的长期命中率，而不是单周盈亏。

其二，「概率→交易结构」正在成为 agent 的独立工种。 礼来案例里，模型没有推荐任何交易；Claude Code 的贡献是把概率报告拆成「报表线」与「underlying 基准」两层，再组装成一条对冲性的双腿结构。机制上，这补上了预测产品最难变现的一环：概率本身卖不出价，把概率翻译成可执行交易结构的解释层才卖得出价。启发：做预测/研究类产品，值得把「谁来把结论变成动作」设计进产品；未来几周值得观察这类两腿结构会不会成为 agentic trading 的标准打法。

其三，公开留痕系列化是小团队最便宜的信任基建（系列第四篇，与前文互为印证）。 从首周股票实验、历史情景回测、中期选举赛前留档到本周的周预测，作者始终用「真金白银+事前钉死+定期结算」积累可信度。启发不变：track record 只能靠事前注册的记录积累；对读者而言这类灰色信源可信度高于匿名总结帖，但仍须记住作者既是裁判（模型方）又是运动员（下注方）。

不确定项：SPY 777 一注的进场方向与价位（转写「6, 7, 65」含混，正文按长射推断标注）；「Brent price on October 20th」疑为 10 月 30 日；布伦特 112 一注截至视频录制未结算；14 个事件的完整清单以转写为准，个别事件（霍尔木兹海峡航运、VTI weekly touch）细节未展开。

ASR 校正表

转写原文（修正版）

[0s - 120s] Hello, hope you are doing well. So I wanted to kind of kick off my series here, but I wanted to do more weekly AI predictions using kind of my Sift Intelligence system, and I wanted to kick this week off by taking like fourteen events, basically just some random events in kind of geopolitics, finance, or elections on Polymarket and Kalshi, and see if we can just do something in this very short span from Monday to Friday, and if we add all of these up, let’s see if we can make any profits. So basically, this is just a test for me, but why don’t I just record it and kind of share the results and see if we can improve over time as my system gets better, hopefully. So yeah, that’s basically what I wanted to do, and let’s just jump straight into the first prediction. I think I’m not gonna go through every one like live, but we’re gonna sum them up. So this is Monday, October 26th, and I guess on Friday we will have have all results Friday evening, I guess. So yeah, let’s just do the first one and see where this takes us. So I actually did get started. So you can see the first one is the US Q T GDP growth, and if you just head back here now, you can see I started running this forecast here, and we are about like 58 percent. We are still doing research on this. Ah, this is the Nvidia finish the the week above this, and also gold price on October 30th. So for the other ones, there’s not much of a hurry, right? This is like the ECB October decision, BOJ October rate decision, shipping targets in Strait of Hormuz, Amazon beats earnings expectations, Eli Lilly beats those, Eurozone GDP, German GDP. Ah, yeah, we talked about this by Nvidia, VTI weekly touch, and Brent price on October 20th, 10 year yield on Friday, and the gold price we talked about. So these are kind of the fourteen events we’re gonna try to predict and see if we can actually can use AI.

[120s - 240s] To do some very short-term prediction in this case, so I thought this could be like an interesting series, like I mentioned, and just to see what we can do with this. So, yeah, I’m just gonna run this now, and I’m gonna take you back when I have all the predictions or forecasts. And of course, on Thursday, Friday, we’re gonna look at the results. I just think this could be like an interesting ongoing series, and of course, I’m gonna bet some money on this. I’m gonna put like ten bucks on each, and we can kind of do the, yeah, the accounting or like see how we are doing over time and plot the results, and we can see if we made money or if we lose this money on this. I think it’s just like an interesting thing if we can see if this system actually work using AI on this and my Sift Intelligence startup model here. You can see we have the GDP. Okay, so US GDP growth Q3. So the market is at 63, and we have like 54%. So I think this could be like an interesting no, maybe so over three percent, maybe do no, since we are like eight gap from the market here. So let me just find this US GDP growth on Polymarket, and let’s place like ten bucks or something on no here. So if you go to my Polymarket account now, you can see if you scroll down here, will GDP or US GDP growth in Q3? So we just placed like ten dollars on this. You can see the dipped a bit, but I don’t know what happened here. But basically, it’s probably gonna go up again. But we placed like ten dollars on this, and let’s see on Friday where this goes. So I thought that just could be interesting. I also have some other bets I’ve been doing lately that is pretty going pretty well, right? So yeah, but I’m gonna talk about that in another video. Okay, so you can see we have the Nvidia results here. So our system gives 66%. Yes, Nvidia finishes above 210. The market is down at 47.

[240s - 360s] But if you kind of look at the market, it is. If you kind of look at the two ten, it says forty seven, but the spread here is like thirty three cents. So this is in kind of like a difficult spot now. So I might hold off a bit just to see where the price develops. And gold is also done. Here we are kind of close to the market. So we end up around four thousand to forty forty around the week. So if we head to calculate, that should put us around this, above this, right? So looks like we get like fifty one here. Let’s say I think that’s fine. So let’s just place something on fifty one. Okay, so it looks like we’re a bit slow here, but let’s just enter at fifty two. Let’s just do like eight bucks or something on fifty two, and yeah, let’s just see on Friday how that goes. Okay, so now I have all the forecasts we’re gonna do for this week. Um, I did this yesterday, and I also placed some positions yesterday. But now it’s the next day when I’m recording this. But basically, we have some few interesting results. So for example, the ECB and the BOJ rate decisions—they are pretty much like dead, right? Eighty eight and ninety five. We do actually disparage like ten percent here, but yeah, I think those are just a skip. So that will be interesting to see. Also on the Eli Lilly beats quarterly earnings at seven forty nine per share. We are on leading more no than the market. So we’re probably gonna we did enter that. And on Brent oil, we kind of came in like a bit more expensive than the market. Same on the yield. Gold, we did cover right, and I don’t think there were any other like big discrepancies on Amazon earnings. We kind of came in exactly the same, right? So yeah, I think we’re just gonna wait for that now. So for you, that’s gonna be like one second, but for me, that’s gonna be like four days or something.

[360s - 480s] So yeah, that is kind of the setup we have now before Friday, and I will take you straight into all the results and let’s see how we did this week. Okay, so we are back. So now it’s been like it’s now Saturday, thirty first of October, right? And yeah, I guess we have all the results from week forty four. So we kind of set out to do all these questions, but of course, I didn’t really get onto the market for the prices we wanted for all of this. But I think we got in at on nine or something. So yeah, I think we’re just gonna check it out how it went and drum roll, please. I guess and boom. Okay, we didn’t do too bad. We got in on nine of fourteen. We did make a very small profit, but there are some interesting things here I want to go through. So you can see just straight away we did actually lose on the GDP, but like I said, we did pick like the nose side here. I know we predicted yes, but we felt kind of we wanted to do a nose side here to have a bigger upside, but that didn’t go in. And on the Eli Lilly beats nine seven forty nine EPS, that came in. We went on the nose side, so that was pretty good. I have an interesting study on the Eli Lilly that I’m going to show you at the end of the video, so you don’t want to miss that. But this I think was our best one maybe. So Eurozone QD Q three GDP between zero point eight and one point one on no, that came in, and this gave us like we got in at thirty five, we paid out hundred. So I think that was our best one. So let’s just look at the development here, right? Remember, we were in this bracket where we we were on the nose side, and you can see all.

[480s - 600s] All the way up here, it wasn’t like 70, right? But yeah, it resolved to no, so that was tick out was our best. So that was a really good prediction by our model. Germany also was pretty good. We came in at no here. We also made like double our money on that, so that was pretty good. SPY hits 77, 77 no, that has to be a bit unlucky, to be honest. I think I’m going to show you that. So basically, you can see we got in around 6, 7, 65 or something, and it barely touched 770, like just up on 77. It never went higher, like all week. I guess it went 771, but after that it just slid down all the way to 756. So a bit unlucky on that, maybe. But yeah, that’s just the way it is. Oil dips to 100. That was in very early and Brent above 112. That is also in. It isn’t really resolved yet, but yeah, the 10 year above 497 that did not go in, and also gold above 4, 000 also didn’t come in. So overall, we were slightly positive. So our PNL is like two dollars. So not the biggest return, but I think we had some interesting good picks and some more questionable. I know this one was a bit of a long shot, but I would say I’m pretty positive because this is our first week doing this, and I’m going to be back next week with a new one with new numbers and stuff like that. Let me know in the comments if you want me to split it up so you can follow along and you can also look at some positions. Maybe you want to enter. Let me know in the comments if you want that. But all in all, I’m pretty happy.

[600s - 720s] Actually, we didn’t lose anything, but we didn’t gain anything either. So yeah, pretty interesting. But I really wanted to to talk a bit about Eli Lilly because we have like a we had like a really interesting position bet around the earnings call this week for Eli Lilly. Remember, here we did no on the seven forty nine EPS. So I just wanted to show you a bit of an interesting trade I did this week with this prediction. So I kind of used Sift Intelligence as my starting point on the. I guess I can bring it up here. Okay, so if you remember, this was kind of the report. So will Eli Lilly beat quarterly earnings? We landed at thirty three, like minus eleven below the market. So a really good prediction, actually. I kind of liked our prediction here. So the outcome was no, so they didn’t. But I wanted to show you an interesting trade I did this week based on the report I got from Sift Intelligence here. So I think I’m just gonna play this video now, and I’m gonna take you back to just finish the video after. Yeah, I explain that. Two days before Eli Lilly reported earnings, Sift made one forecast. Read it closely, and it points to two trades: bet against the beat on Polymarket and buy the stock. Polymarket asked, “Will Lilly’s earnings per share beat seven dollars forty nine? " The market said forty four percent. Sift said thirty three. Three forecasts, all within five points. The first clue is the line itself. Lilly had already announced a one off charge of four dollars fifty seven a share, and it sits inside that number. So the real bar is about twelve dollars, too high. Trade one, no on Polymarket. The second clue is Wall Street.

[720s - 836s] Analyst estimates ran as low as five dollars thirty-eight because many hadn’t fully counted that charge, and Lilly had beaten underlying estimates by about twenty percent, two quarters running. So even a number just under the line looks like a beat. Trade two, long Lilly on Interactive Brokers. This is where an AI agent comes in. Claude code running Opus 5.5 reads the Sift forecast and asks two questions: Will they clear the line? No. Will they beat Wall Street by more than the options market has priced in? Yes. Two answers, two trades. On October twenty ninth, Lilly reported seven dollars forty-eight, one cent under the line. The no paid out plus seventy-five percent, and the stock, it jumped about eight percent. Sales came in far above forecasts, and Lilly raised its outlook. One forecast, two trades that look like opposites, and both were right. Sift gives the probability. The agent finds the trades. So yeah, that was pretty interesting, right? So I found something I haven’t done before, and I think these kind of opportunities is something you can do when you have this starting point, as we did with the Eli Lilly quarterly earnings report here. But yeah, basically that was week one, I think, and we didn’t do too bad. We got like a small profit, like almost like two dollars. But all in all, I really enjoyed it, and if you enjoyed it, give this video a like, and if you want me to do it like in two videos, let me know in the comments. So yeah, thank you for tuning in. Hopefully we can make this into a series and see if there is some interesting things we can do with AI and predictions on these short time horizons. So yeah, I hope you enjoyed it, and hopefully I’ll see you again very soon. Have a nice day.