Leigh Drogen（Starkiller Capital）解密群眾外包機器學習打造 StatArb 金庫

0:00開場與CrowdSent介紹

0:00我非常興奮，因為等一下我們要在這裡發表一個全新的東西。為了讓大家有更好的背景，我先邀請我們的合作夥伴 CrowdSent 的創辦人 Jason Rosenfeld 上台，來說明他們這幾年在做什麼。I'm incredibly excited to be launching something here in a couple of minutes, but in order to set the stage for that, I'm going to have our partner at CrowdSent, the founder, Jason Rosenfeld, come up and explain what they've been working on for the last couple of years.

0:24謝謝 Leigh。我會先講幾分鐘，等一下 Leigh 也會再上來。我們很高興跟大家介紹 CrowdSent，這是一個群眾外包的機器學習挑戰，未來將驅動 Hyperliquid 上的 StatArb 市場中性金庫。簡單來說，每天都有資料科學家來我們的平台，針對每個 Hyperliquid 永續合約預測未來表現，然後我們把所有人的預測彙整成一個 整合模型，最後把這個 整合模型 轉成可以交易的投資組合。Awesome. Thanks, Leigh. I'll talk for a few minutes, and then Leigh will come back up as well. We're very excited to tell you about CrowdSent, which is the crowdsourced machine learning challenge that's soon going to be powering a StatArb market neutral vault on Hyperliquid. So in sum, we have data scientists come to our platform every day. They're ranking every hyperliquid perpetual based on how they think future performance is going to do. Then we aggregate all of those predictions into a single meta model and then turn that meta model into a portfolio that can be traded.

1:08StatArb 與傳統避險基金模型

1:08我想先了解一下現場的觀眾，既然你們會來聽這場標題有 StatArb 的演講，有多少人知道什麼是 StatArb？好像只有少數人。其實這不難懂，我們就是用機器學習和統計方法，找出市場的低效率並加以利用。因為我們是金融避險基金產業，所以這裡有一大堆免責聲明，不過我就不一一唸了，主要是這不是證券發行，有興趣可以上我們官網看完整聲明。I'm just curious to get a sense of the audience, your watching a talk that has StatArb in the title. How many people know what StatArb is or are familiar with it? Okay, so a few people. Good news is it's not too complicated. We use machine learning, we use statistics to try to identify inefficiencies in the market and then exploit that. And so because we are in the finance hedge fund industry, I have a nice wall of text of disclaimers here for you. I'm not going to read the whole thing. Mainly, this is not a securities offering. You can go to our website for more disclaimers and disclosures.

1:51CrowdSent 的核心理念，就是把傳統避險基金的運作模式顛倒過來。傳統基金通常只會聘請少數幾位投資經理或研究員，把他們關在房間裡秘密工作。如果你不搬去紐約、倫敦或新加坡，通常很難成為這些『聰明人』之一來參與決策。但我們認為，人才是全球分布的，機會卻不是，所以我們想讓每個人都有機會。And so, with that out of the way, I'll tell you a bit about the premise of CrowdSent. And the premise is to take the traditional hedge fund industry model and then flip it on its head. So what is the traditional fund model? It is basically hiring a few portfolio managers or researchers, hiding them in a room and having them do their work in secret. Usually, if you're not moving to New York City, London, or Singapore, you're not going to be one of these few quote-unquote smart people in the room making these investment decisions. And so at CrowdSent, we think there is an alternative, better model. Basically, talent is globally distributed, but we've seen that opportunity is not. So let's give an opportunity to everyone.

2:38開放平台與群眾智慧

2:38我們打造一個開放平台，任何人都可以下載資料，甚至帶自己的資料來，然後提交自己的模型。我們會把所有模型的預測彙整起來，並且只用實際市場資料來評分，所以很清楚誰表現好、誰不行。問題是，如果激勵機制設計得好，這種新型避險基金模式能不能跟 Citadel、Millennium 這種傳統巨頭競爭？我們認為答案是肯定的，因為群眾外包機器學習其實已經有十幾年歷史了。If we have an open platform where anyone can download the data, they can bring their own data and then submit a model. You know, we see that we can clearly aggregate all of those. And then the models are only scored on live data, so we know who is doing well and who isn't. And so the question is, with properly aligning incentives, can you build this alternative hedge fund model that can compete with the likes of Citadel and Millennium? Obviously, we think the answer is yes, and there's some precedent for this. Crowdsource machine learning already works. It's been out there for 10, 15 years.

3:16現場再問一下，有沒有人聽過 Kaggle 這個資料科學競賽？好像有一些人。那更貼近加密圈，有沒有人聽過 Numerai？Numerai 是做股票市場預測的群眾外包避險基金，跟 CrowdSent 很像，只是我們是做 Hyperliquid 永續合約。其實我們 CrowdSent 團隊本身也有參加 Numerai，累積了多年經驗，知道怎麼辦比賽、怎麼做出好的機器學習模型。Maybe another poll of the audience. I'm curious if people have heard of the data science competition called Kaggle. Okay, a few people. And then more crypto-friendly, how many people have heard of Numerai? Okay. So Numerai runs a hedge fund. They run an equity hedge fund. It's a stock market prediction, similar to what CrowdSent is doing, but CrowdSent is doing it with hyperliquid perpetuals. But Numerai works, and CrowdSent actually participates as machine learning engineers to Numerai. And so we've competed, we've staked for a number of years now, and we know what it takes to build a tournament, and we know what it takes to build good machine learning models.

4:01模型彙整與投資組合建構

4:01這裡有一個流程圖，左邊是各個參賽者的模型排名，每一欄都是一個人的預測。例如最左邊那位認為 Athena 未來 30 天表現最好，Dogecoin 最差。每一欄都是不同人的預測，然後我們每天把所有預測彙整成 整合模型。接著我們把 整合模型 轉成 多空 投資組合，比如做多前 40 名、做空後 40 名，這樣就能組出一個市場中性投組。這裡面有很多組合改善和中性化的技巧，等一下 Leigh 會再多說。So here's a nice visual of kind of what the process looks like. So on the left, we have a bunch of model rankings. Each column that you see is somebody's prediction. So you'll see on the far left, someone thinks Athena is going to do the best over the next 30 days, and at the bottom of the list is Doge. So Doge may do the worst. And then each new column is somebody else's prediction. And then you see the meta model, which is aggregating all of those signals every day. Then we take the meta model column and we turn that into a long-short book. So maybe you want to go long the top 40 of the meta model. Maybe you go short the bottom 40 of the meta model at the same time to create that portfolio. And there are a lot of ways to do this. You can do different portfolio optimization. You can do different neutralization, and we'll see more about that with Leigh. So.

4:57參與方式與AI代理人時代

4:57參加其實很簡單，也越來越有趣。我們有個小遊戲，基本上只要五行 Python 程式碼就能參加。但現在 AI 代理人很流行，大家其實不用自己寫程式，直接問 AI 代理人就好。我們提供免費資料，你可以什麼都不帶來，也可以帶自己的資料，或用自己的訊號。我們有排行榜，大家都在實際市場資料下競爭，氣氛很激烈。最近我們還推出 CrowdSent Cloud，讓你的 AI 代理人可以用外部算力來訓練模型。What does it look like to actually participate? It's very easy. It's also becoming a lot more fun. You may have seen the little clip of our game. It really is just kind of five lines of Python. But if you've been following AI agents, nobody's writing code or Python anymore, so it's really just asking an AI agent. We give out free data, so you can just come with nothing. Or you can bring your own data, or Thagno Signal is always good. We have a leaderboard that's scoring everybody out of sample, so it's pretty competitive. Everyone wants to do pretty well. And we recently launched CrowdSent Cloud, which is external compute for your AI agents to go build your machine learning models.

5:41為什麼這是個轉捩點？因為過去半年到一年，最強的預測者已經是機器，最會訓練模型的也是機器。現在做機器學習模型比以前容易又有趣。即使你不懂機器學習，也歡迎來 CrowdSent 跟我們的 AI 代理人互動，說不定你不小心就變成機器學習高手。我們看到 bot 拿下 Metaculus Cup 前兩名，這真的正在發生。我們的代理人叫 Centaur，歡迎大家來試試看。And so why is this kind of the inflection point? It's because over the last six months to 12 months, the best forecaster is now a machine. The best machine learning model builder is also now a machine. So like I said, it has never been easier. It has never been more fun to kind of build these machine learning models. Even if you don't know machine learning, I encourage you to go to CrowdSent. You can start talking to our agent, and you may accidentally become a machine learning expert. We saw that bots took first and second in the Metaculus Cup, so this is really, really happening. Our agent is named Centaur, and again, yeah, I just encourage everybody to check it out.

6:27從訊號到實盤：與Starkiller Capital合作

6:27我剛剛講了很多關於挑戰賽和訊號產生，但要把這些訊號變成實際可交易的投組，還有很多工作要做。這也是我們為什麼要跟 Leigh Drogen 和 Starkiller Capital 合作。現在請 Leigh 上台，謝謝大家。So I've talked a lot about the challenge side of things and what we've built to build the signal, but there's a lot of work that goes and moving the signal to a live portfolio. So that's why we partnered with Leigh Drogen, Starkiller Capital. Do you want to come on up and tell them about it? Thank you, Jason.

6:45什麼是 Starkiller Capital？Starkiller 是我在 2021 年創立的傳統機構型避險基金，目前有兩個策略：一個是量化驅動的流動型代幣基金，主打趨勢追蹤和動能模型，還有一些量化基本面因子；另一個是 DeFi 市場中性收益基金，也會做一些套利。根據 Crypto Insights Group 的資料，這兩個基金的表現都在同業前 10%。我們對這些成果很自豪。So what is Starkiller Capital? Starkiller is a classic institutional hedge fund that I founded back in 2021. We run two different strategies today, two different books. One is a quant-driven liquid token fund, really based off of trend following and momentum models with a bunch of quantitative fundamental overlay. The other one is a DeFi market neutral yield fund that also does some arbitragey type stuff. Both funds in the top decile relative to their peers, according to Crypto Insights Group. So we're really proud of what we built.

7:23群眾外包與區塊鏈新價值分配

7:23我的背景是從股票 stat arb 圈子開始的，十多年前我和 Jason 就認識了，當時他在 Millennium 做研究員，用的是我創辦的 Estimize 公司提供的資料。Estimize 是全球最大的股票基本面群眾外包平台，跟 Thomson Reuters、Bloomberg 等傳統機構競爭，但我們的預估來自 buy side，而不是 sell side。當時我們很困擾，因為沒辦法把價值回饋給貢獻者，因為 2011 年還沒有區塊鏈和代幣經濟。現在有了區塊鏈，我們終於可以把價值分配給參與者。My background begins in the equity stat arb community, which is actually where I met Jason over a decade ago when he was a researcher at Millennium, using data built by a company that I founded called Estimize back in 2011. Estimize was a big crowdsourcing competition for not machine learning models, but equity fundamental estimates, so earnings per share, revenue, EBITDA margins. Estimize eventually ended up becoming the largest crowdsourcing platform for equity fundamentals in the world, to rival Thomson Reuters Ibis or First Call or Bloomberg estimates, but instead of from the sell side, the estimates were from the buy side. But one thing that we were always a little annoyed about was that we were not really able to give a ton of monetary value back to our contributors because there was no mechanism back in 2011. Now we have decentralized blockchains and the ability to imbue tokens with the value from certain businesses,

8:39六個月前我們開始分析 CrowdSent 的資料，發現 整合模型 的準確度非常驚人。84% 的天數 整合模型 都比中位數模型更準，平均每日分位數在第 69 百分位。這個競賽產生的資料品質一看就知道很高。which brings us to six months ago when Starkiller started to look at the CrowdSent data, and what we found was incredible. The meta model associated with the CrowdSent data happens to be incredibly accurate. 84% of days it's more accurate than the median model. The average daily percentile in the 69th percent. So the very basics of, is this competition actually producing quality data? was very obvious right away.

9:20從模型到可配置投資組合

9:20我們先用一個很簡單的模型來測試，把 整合模型 直接做成等權重的 多空 投資組合，沒有做任何因子中性化，結果已經有很穩定、很高的夏普比，尤其跟比特幣相比更明顯。我們相信這個流程可以複製到很多不同資產類別：先辦挑戰賽，產生 整合模型，再轉成風險管理的投資組合，讓個人和機構都能參與配置。So what we did was we looked at a very naive model, actually turning that into a portfolio. So this is an equally weighted book, long and short. Jason alluded to this, and even that equally weighted book, not neutralized for any factors, already produces very consistent, high sharp returns, especially relative to Bitcoin. So what we're excited about here is there is a process that we believe is repeatable across potentially many different asset classes and assets to take a challenge, turn it into a meta model, and then turn it into an actual risk-managed portfolio that could be allocated to by both individuals and institutions.

10:18今天我們就是要正式啟動這個金庫的建構流程，這個金庫很快就會上線，等一下會再說明細節。這個金庫就是用 Starkiller 幫忙打造的模型來實際交易。And so what we're here to do today is actually launch the process of building a vault, which will be available soon, and we'll get to that in a second. That actually trades the model that Starkiller has helped build here.

10:36投資組合改善與市場中性

10:36要把 整合模型 變成真正可用的投資組合，還有很多細節要處理。首先要做因子中性化。在加密貨幣市場，沒有像股票市場那樣明確的產業分類，所以我們用純統計方法來中性化，這對機構資金來說非常重要。我們也會根據參賽者模型的信心水準來加權，信心越高的模型在投組裡權重越大，這樣可以進一步提升表現。So when we look at actually building a portfolio off of that meta model, there are a whole bunch of things that need to be done. One is we need to neutralize all of the different factors. Now in crypto, we don't have a lot of classic factors like we have in the equity market. There isn't really a good agreed-upon sector classification, and so these factors that we use to neutralize the portfolio are purely statistical, and they add a ton of value in producing a book that can be allocated to institutionally. We also conviction weight the portfolio, so obviously the competitors in the tournament produce models that have more conviction in the top decile of the universe than the, you know, the third decile. And when we do this in different ways, it obviously improves the portfolio.

11:43最後我們得到的投資組合，年化報酬率高達 20% 以上，夏普比接近 2，波動率 9%，最大回撤僅 7%。這些都是非常優秀的 stat arb 策略指標。最讓我們興奮的是，這是一個專屬的資料集，短期內不會被市場套利掉。And what we eventually get is a portfolio that produces high 20% returns with a, you know, almost two Sharpe, nine vol, and a max 7% drawdown. Those are really high quality statistics associated with a stat arb strategy. And the thing that we're excited the most about here is that this is a proprietary data set that nobody else has, that should not be arbitraged, the value of which should not be arbitraged away anytime soon.

12:19金庫啟動與未來展望

12:19我們今天正式開放這個金庫的報名，這是一個有容量限制的策略。大家現在就可以到 crowdsent.com/challenge 登記，金庫會在幾週內上線，我們非常期待。So what are we doing here? We're launching. The sign up for the vault. This is a capacity constrained strategy. You can go to crowdsent.com/challenge today and get on that list. The vault will be launching in a number of weeks here. We're incredibly excited for it.

12:44最後我想談談長遠的趨勢。隨著 AI 模型崛起，我們相信大型避險基金在執行系統化策略上的優勢會被逐漸侵蝕，模型建構能力正在民主化。我們認為，未來 alpha 的來源只剩兩件事：一是難以量化的勤勉策略，二是你是否擁有獨特的資料集。其他都會變成競爭激烈的紅海。當個人也能用 AI 參加 CrowdSent 這類競賽，這會成為價值捕捉的新管道，無論是個人還是機構都能透過比賽獲得回報。我們非常期待這個未來，有興趣深入討論的歡迎找我或 Jason 聊聊，謝謝大家！The last thing I wanted to talk about is, over the long term, what's going on with the advent of these AI models and the fact that our belief in getting involved in this project, you know, part of it is that. As AI models really eat away at the advantage that large hedge funds have in executing on these systematic strategies, they are definitely democratizing the ability to build models. Our belief is that the future of alpha in almost all markets is going to come down to two things. Diligence in hard-to-quantify strategies, and do you have a unique data set that nobody else has? It is going to be a treadmill for everything else besides those two things. And as individuals get their hands on these AI models and point them at competitions like CrowdSent, our general belief is that this can be a sync for value capture, where an individual or an institution produces models, they participate in tournaments like this, and they receive value back through that competition in different ways. And so we're incredibly excited. Please get in touch with Jason or I if you'd like to discuss those at more length. Thank you.