What elite analysts actually do — in their own words and work products — and how AI is changing it
Across firsthand accounts, the institutional process runs: idea sourcing (screens, sector coverage, network) → business analysis and a driver-based model → deep dive on the 2–4 factors that actually move the stock → bull/base/bear case construction with risk/reward asymmetry → pitch to the PM → monitoring against catalysts. James Valentine — whose Best Practices for Equity Research Analysts the CFA Institute adopted into its curriculum — frames the core skill as identifying the critical factors per stock and building information networks that give proprietary insight on exactly those (free sample chapter).
Paul Enright, who ran consumer and tech books at Viking Global for 12 years, gives the best single public account of the elite seat in his "Buy Side Primer" (Invest Like the Best, transcript): he separates "great researcher" from "great stock picker," describes what good versus great digging looks like (primary research that tests the thesis, not confirms it), and stresses that business analysis is not stock picking — the stock call is a separate judgment about expectations and timing.
Practitioner threads and accounts converge on the same philosophy. From a top-voted long/short analyst on WSO:
"You are not trying to out-model the CFO… assume guidance embeds the base case, then find key points that have changed since guidance… The goal is to take a reasonable over/under. It's about taking a view on the company, not the model."
In another thread, the consensus: public-side models are far less granular than PE models because the data doesn't exist — and the edge only requires a different view on one assumption the market has wrong (e.g., net retention of 115% vs. street's 110%). Even a PE principal concedes "the idea that 'the model' helps you make good investments… is a fallacy."
To see what real buy-side-grade models look like, study published ones: Abdullah Al-Rezwan (MBI Deep Dives, ex-buy-side) publishes full deep dives with KPI-driver operating models — see his free Autodesk and Otis deep dives — and his research process post describes the actual sequence: read the filings → build the model shell → write a question list → calls with IR and bulls/bears → write to think. Alta Fox Capital publishes its actual investment decks (XPEL, REVG, Enlabs…) plus the "Makings of a MultiBagger" study it uses as a sourcing framework.
The depth of work depends on the seat. Single managers (Viking, Maverick, Greenlight lineage) underwrite multi-quarter intrinsic-value gaps with deep primary research. Lee Ainslie (Graham & Doddsville interview) says "the most critical factor we're trying to evaluate is the quality of management," and explicitly warns against valuation-only investing. Larry Robbins (G&D interview) describes Glenview's edge as exhaustive, consistent coverage of a defined universe — "consistency of effort wins the game." At multi-managers (Millennium: 330+ pods, ~$79bn; Point72: 185+ teams), the game is getting quarters right: quarterly projections, beat/miss handicapping, catalyst-linked positioning inside tight risk limits (M&I on multi-managers). Pod analysts re-underwrite assumptions intra-quarter; the model doubles as a sentiment and positioning tool.
Multiples as shorthand: the working valuation in most pitches is earnings power 1–3 years out times a defended multiple range under bull/base/bear scenarios — visible in the real decks in Section 6 (Greenlight's Sohn pitches, Alta Fox's decks, Kerrisdale's short reports all follow this pattern).
Reverse DCF / expectations reading: start from today's price and solve for what it implies. Speedwell Research's explainer shows their actual Meta (Jan 2023) reverse DCF — the market was pricing Meta as if growth was over. The formalized discipline is Rappaport & Mauboussin's Expectations Investing (free tutorials): read price-implied expectations, judge revision likelihood with competitive analysis, express it in probability-weighted scenarios anchored on base rates.
Asymmetry as the filter: PMs look for "60/40 and 70/30 coins," not 50/50 bets (Fundamental Edge), with ~3:1 upside/downside commonly cited as the conviction bar. Position sizing follows fractional-Kelly logic — Thorp's Understanding the Kelly Criterion (PDF) is the primary source — but at small fractions: full Kelly carries ~1-in-3 odds of a 50% drawdown.
The fastest way to absorb elite practice is reading analysts and PMs describing it directly. The highest-signal primary accounts found:
| Who | Key process insight | Source |
|---|---|---|
| Paul Enright (Viking Global, 12 yrs) | "Buy Side Primer": researcher vs. stock picker; what great digging looks like; building blocks of an effective pitch | ILTB transcript |
| Lee Ainslie (Maverick) | Management quality as the central evaluation; longs ~2x shorts; sector heads own deep verticals; valuation alone is not a thesis | G&D interview |
| Larry Robbins (Glenview) | Exhaustive consistent coverage of a defined universe; modeling FCF deployment; consistency over flashes of insight | G&D interview |
| Dan Sundheim (D1; ex-Viking CIO) | Reading-first routine; qualitative judgment over model precision; risk/shorting process post-GameStop | Cheeky Pint interview |
| Li Lu (Himalaya) | Research as investigative journalism — physically verifying assets — vs. the "95% of investors" forecasting quarters behind terminals | 2006 Columbia lecture (PDF) |
| Mohnish Pabrai | Checklist built from autopsies of other investors' documented mistakes; cloning as deliberate sourcing | G&D interview |
| François Rochon (Giverny) | Annual "owner's earnings" scorecard tracking intrinsic value vs. price; public "Mistakes du jour" error audit in every letter | Letters archive |
| John Huber (Saber Capital) | Time-horizon arbitrage as the only durable edge; reading and journaling over screening | Letters & notes |
| Howard Marks (Oaktree) | "Something of Value" (2021): a top practitioner publicly revising his own valuation framework — qualitative depth can justify holding without precise quantification | Memo (PDF) |
Deeper archives: the Graham & Doddsville newsletter (Columbia; every issue since 2006 — interviews with Greenblatt, Ainslie, Robbins, Li Lu and dozens of Tiger-descended PMs asked specifically about sourcing, modeling and diligence; latest issue PDF); Value Investor Insight interviews, which force managers through one idea end-to-end (e.g., Rochon, Turtle Creek); and the canonical books of practitioner interviews: Pedersen's Efficiently Inefficient (verbatim process interviews with Ainslie, Asness, Paulson, Chanos), Schwager's Hedge Fund Market Wizards, and Freeman-Shor's The Art of Execution — an allocator's audit of 45 top managers' actual trades showing execution behavior, not idea quality, separated winners.
The named disciplines that recur in elite practice, each with its primary document:
| Fund | What they've deployed | Source |
|---|---|---|
| Citadel | "Citadel AI Assistant" used daily by nearly all equities investors — scans transcripts/filings, summarizes broker research, flags risks. PMs barred from offloading judgment. Griffin: GenAI "just falls short" for alpha. | Reuters · Bloomberg |
| Balyasny | Internal platform used by ~95% of ~180 teams; agents cut macro scenario analysis from 2 days to ~30 min; proprietary finance-tuned embeddings beat OpenAI's in internal tests. | OpenAI case study · Hedgeweek |
| Viking Global | "VikingGPT 2.0" fields ~3 queries/min, usage +400% YoY; deputy CIO uses it as a "sceptical avatar" to stress-test theses. No investment calls. | Bloomberg |
| Bridgewater | AIA Labs $2bn fund (July 2024): ML as primary decision basis, humans on risk/execution — "unique alpha uncorrelated to what our humans do." | Bloomberg |
| Man Group | "AlphaGPT" generates, codes and backtests signals — several dozen AI-originated signals live (July 2025), human-vetted. | Bloomberg |
| AQR | ML now drives ~one-fifth of signals in the flagship multi-strategy fund; Asness "surrendered to the machines." | Bloomberg |
| NBIM | Uses Claude to screen every company entering the portfolio on day one for ESG/ethical risks. | CNBC |
Adoption data: 95% of hedge fund managers use GenAI (AIMA 2025); 55% of asset managers have AI in at least one investment process but only ~5% grant it any decision authority (Mercer 2026); alternative-data adoption hit ~90% of advisers, led by card-transaction data (Lowenstein Sandler 2025 (PDF)). The strategic divide: Griffin and Asness say efficiency-not-alpha; Bridgewater and Man Group live-trade AI-originated decisions. The market overwhelmingly sits with Griffin for now.
More useful than fund PR: named practitioners publishing exactly how they use LLMs day to day.
Labor-market read from the WSO hedge fund forum: "what used to take 2 analysts can now be done by 1"; expect more top-heavy teams as AI absorbs the modeling grunt work junior seats were built on; lean single managers use AI for breadth ("cover more sectors, peer into foreign markets") rather than cuts. Notably, no major fund's investor letter yet describes its internal AI process in primary terms — funds discuss AI as a position, not a process.
Actual artifacts by top investors — the best study material there is. Every link below was fetched and confirmed live on 8 June 2026.
| Artifact | What it is | Link |
|---|---|---|
| Greenlight — Sohn 2026 deck | Einhorn's actual conference pitch, from Greenlight's own site (2025 and 2024 decks also posted there) | |
| Einhorn — Lehman short speech (2008) | "Accounting Ingenuity" — the famous Sohn speech, from Yale's FCIC archive | |
| Pershing Square — Herbalife deck | "Who Wants to Be a Millionaire?" — the ~334-slide short thesis (Dec 2012) | PDF (archive.org) |
| Pershing Square — 2026 annual presentation | Position-by-position theses on the current portfolio, published by the fund | |
| Starboard — Darden deck (2014) | The ~294-slide "Olive Garden" activist plan — the genre's most famous deck | |
| Kerrisdale — CoreWeave short (2025) | Full short thesis with model assumptions; Kerrisdale publishes all reports free | |
| Muddy Waters — Sportradar (2026) | Forensic short report; full archive on their site | |
| Hindenburg — Adani / Nikola | Full forensic reports, still live | Adani · Nikola |
| Sohn Idea Contest archive | ~20 winning pitch decks 2016–2021, direct PDFs | Archive |
| Alta Fox — research library | A fund publishing its actual investment decks + the "Makings of a MultiBagger" sourcing study | Library |
| Archive | What it is | Link |
|---|---|---|
| Nomad Partnership letters 2001–14 | Sleep & Zakaria's complete letters — the modern classic on long-horizon process; authorized copy from Sleep's own foundation | |
| Buffett Partnership letters 1957–70 | The original — position categories (generals/workouts/controls), sizing logic, expectations-setting | |
| Howard Marks — complete memos 1990–present | Full archive plus a single compiled PDF on Oaktree's site | Archive |
| Third Point quarterly letters | Loeb's letters — thesis summaries per position (Q2 2025 verified) | |
| Michael Burry — Scion letters | Compilation of Scion's 2000s annual letters | |
| Giverny Capital letters 1993–present | Rochon's owner's-earnings scorecard and "Mistakes du jour" every year | Archive |
| Ongoing letter aggregator | Curated quarterly collection across hundreds of funds | FinMasters |
| Source | What it is | Link |
|---|---|---|
| MBI Deep Dives | Ex-buy-side analyst's full deep dives with operating models; Autodesk and Otis dives are free PDFs; model library subscriber-gated | ADSK dive · Models |
| Speedwell Research | Independent shop publishing institutional-length reports (50–170 pp); frameworks and reverse-DCF explainers free, full reports paid | Site |
| ValueInvestorsClub | Archive of member-written long/short write-ups by practicing professionals; free delayed guest access | VIC |
| Graham & Doddsville archive | Every issue since 2006 — the richest free archive of fund analysts describing process verbatim | Archive |
| Damodaran spreadsheets | The one non-practitioner source worth keeping: full valuation model library (ginzu DCF, sector models), free, no signup | Spreadsheets |
Synthesizing the firsthand accounts into a sequence — the elite process with the AI layer where practitioners actually put it: