Coatue - The Investment Decision Function

# Coatue - The Investment Decision Function A hedge fund is a function. It takes external and internal context and returns one tuple: ==what to buy, how much, and when==. - Source: talk by the Head of AI at Coatue ($93bn hedge fund), [@fyxlong](https://x.com/fyxlong). - Core claim: investing is a ==compression problem==. All of the world's information goes in. One decision comes out. - The hard part is not the models. It is the complexity of the graph, and how the org absorbs it. ## 1. The decision function ```mermaid graph LR subgraph EXT[External context] N[News] P[Prices] F[Filings] end subgraph INT[Internal context] POS[Current positions] LIQ[Liquidity + risk limits] CS[Past case studies] H[History with sector / mgmt] end EXT --> FN{{"f(external, internal)"}} INT --> FN FN --> T["(selection, size, timing)"] ``` - Principle: ==everything an investor sees on screen, an agent must see via API==. - External data: news, prices, filings. All via API. - Internal data: positions, liquidity and risk limits, past case studies, history with a sector or management team. - Coatue's edge: decades of rigorous internal documentation. The internal context already exists as data. - The reality: you cannot dump all of it into a [[Context Window]] and get "three magic numbers". It is not that easy. ## 2. The over-simplified version ```mermaid graph LR G[1. Gather internal + external] -->|output = input| A[2. Analyze] A -->|output = input| D[3. Decide] G -.- g1[dozens of sub-steps] A -.- a1[dozens of sub-steps] D -.- d1[dozens of sub-steps] ``` - Each step has an input and an output. - The output of one step is the input of the next. - Each step has dozens of sub-steps. Each sub-step needs the best tool for the job. ## 3. Pick the right runner per step A "runner" is who or what does a step. Four options, in order of the complexity they can model: | Runner | Strength | Example | |---|---|---| | Deterministic code | Fast, reliable, cheap | `if PE < 15: do X` | | Classic ML | More complex relationships than code | Factor or pattern models | | LLM | Even more complex relationships | Summarize an earnings transcript | | Person | Judgment, trust, relationships | Meet the management team | ```mermaid graph TD S[Sub-step] --> Q1{Rule is explicit?} Q1 -->|Yes| C[Deterministic code] Q1 -->|No| Q2{Structured data, learnable pattern?} Q2 -->|Yes| ML[Classic ML] Q2 -->|No| Q3{Language / unstructured, no relationship at stake?} Q3 -->|Yes| L[LLM] Q3 -->|No| HU[Person] ``` %%The decision tree is my reading of the talk, not his slide.%% - Goal: ==maximize absolute quality per step==. Not "as many steps LLM'd as possible". - "There is no glory in automating things for the sake of automating things." - People still matter. You do not send an LLM, or an `if` statement, into a management meeting. See [[Human-in-the-Loop Systems]]. ## 4. The realistic version ```mermaid graph LR IN[Inputs] --> S1[Gather] S1 --> B1{Branch} B1 -->|path A| S2a[Analyze A] B1 -->|path B| S2b[Analyze B] S2a --> B2{Branch} S2b --> S3[Cross-check] B2 -->|path C| S3 B2 -->|path D| S4[Deep dive] S4 --> S3 S3 --> OUT[Decision] S2a -.- K["Knobs per node: runner(s), model size, thinking effort, hooks, skills, ..."] ``` - The real function is not linear. It has branching logic and many forks. - Each node has an optimal runner. In practice, often several runners per node. - Each node also has more knobs: model size, thinking effort, hooks, [[Agent Skills as Codified Domain Expertise|skills]], and more. - Each data source has its own chain of steps. The firm-wide graph is "massively complex". ## 5. The search space is basically infinite $ \text{configurations} \geq 20^{10 \times n} \quad (n = \text{nodes}) $ - His estimate: at least 10 knobs per node, at least 20 valid values per knob. - One node alone gives $20^{10} \approx 10^{13}$ options. - You cannot search this space by trial and error. You need to cut it into parts. > [!question] The key question > "What can I actually compartmentalize in a way that we can understand and evaluate?" - Compartmentalize the graph into sub-graphs you can evaluate one by one. - Each part needs a North Star metric to hill-climb. Without one, "you're just kind of… fidgeting around." See [[Evals]] and [[Domain Experts as Eval Builders]]. - The real challenge: lay the graph out so people understand it technically ==and absorb it culturally==. "That is the great challenge of applying AI in investing." ## 6. Beyond investing - The general form: a ==state machine that takes in the world's information and understands it well enough to make coherent decisions==. - The same pattern applies to policy, nonprofits, drug discovery, and corporate strategy. > [!note] Link to Bridgewater > [[Bridgewater PAT - Compiling Research into Deterministic Agents]] is a worked example of the same idea. Bridgewater compiles analyst workflows into deterministic code where it can, and uses LLMs only where it must. ---- tags: #boat #ai #investing links: [[AI native hedge fund moc]] [[Bridgewater PAT - Compiling Research into Deterministic Agents]] [[Evals]] [[Human-in-the-Loop Systems]] [[Context Window]] [[Separate the Decision from the Implementation]]