One agentic workflow now does 1,000 hours of hedge fund analyst work.
Aakarsh Ramchandi founded the data team @ Third Point, built screening engines @ FactSet, & now builds agentic research tools @ RavenPack.
"There's gonna be a full convergence of quant and qual. Most discretionary analysts I know are somewhere in their Claude journey β and the quants are going the other way around."
We cover:
- Year one at Third Point: onboarding 100 data sets with a team of 4 β & why they kept point-in-time copies of every vendor feed to catch panels that silently changed overnight
- The Dan Loeb pitch story β a 45-page deck, six weeks of work, he stops at page 26, asks one question, & the whole thesis breaks
- "Kind but not nice" β the zero-politics office where everyone gets corrected by elite people daily
- Why analysts don't want your forecast β they want facts in Excel, red-green-blue, formatted their way
- Hedging a concentrated activist book with alt-data short baskets built from a 400-500 factor model
- Why Nvidia broke the Barra model β & building custom semiconductor factors instead
- The agentic earnings preview: 8-9 step workflows, 35M tokens per run, ~1,000 hours of analyst work encoded
- Self-improving loops β agents reviewing their own last 10 traces & patching their mistakes
- The WorldQuant hackathon: 7,000 quants turning unstructured text into 35M unique time series
Highlights:
(00:00) Intro
(01:38) Founding Third Point's data team in 2017
(03:55) Six months building point-in-time data infrastructure
(06:20) How an event-driven fund actually uses alt data
(12:40) Team structure & the original forward deployed engineer
(17:10) Nobody wants your forecast β just give it to them in Excel
(19:35) Measuring signals: direction, point estimates & confidence intervals
(24:05) Working with Dan Loeb β the elite bullshit detector
(26:05) The page-26 "Why?" story
(28:55) 5AM Saturdays & discipline that compounds
(32:05) Kind but not nice: the zero-politics office
(33:55) How an activist creates alpha by re-running the business
(43:10) Hedging the book with alt-data short baskets
(50:40) Why Nvidia broke standard factor models
(56:25) From search to RAG to agents
(1:04:20) Opus 4.5 changes the game: 70% β 90% accuracy
(1:11:00) Anatomy of an agentic earnings preview β 35M tokens per run
(1:17:20) Ambient agents: the always-on Jarvis
(1:19:40) Self-improving loops & encoded judgment
(1:20:20) Finance in 10 years: the full convergence of quant & qual