An Investment Fund as a Function

Key claims

• The function: investment_fund(information) → decision, where information is news, filings, prices, data and risk limits, and the decision is one tuple: stock selection, size, time.
• You can’t just ask a model. The problem has to be broken into sub-functions you can evaluate, like a series of lossy projections into lower-dimensional space.
• Many valid decompositions: start from stocks and their KPIs and find the themes that move them, or start from themes and find the most exposed stocks.
• And many ways to run each step: one agent per KPI vs. per stock; one reviewer vs. ten reviewers on different models with a majority vote, possibly weighted by model size and thinking time.
• The graph keeps moving: the optimal shape shifts with each new model release.
• Performance over AI usage: often deterministic code, statistical methods or humans are the best implementation of a step.

Notable replies

• Sizing came up as the hardest part with noisy information; Long: “everything has been pretty darn hard.” Others questioned whether a static graph can handle non-stationary markets.

Connections

• AI Agents · Investing Process · Coatue