Bootcamp project · Vector Institute

Agentic Forecasting Live

A two-part series describing work that Ethan did as a personal bootcamp project in preparation to build a live forecasting benchmark.

This is a fork. The library, reference implementations, and strategy guides come from VectorInstitute/agentic-forecasting. This repository is a worked example of what a participant project can look like: take the foundation, pick a market, run a leak-minimized evaluation of agents that read the news, and write up what survived scrutiny.

• Part 1 of 2

  An honest scoreboard for market forecasting

  A frozen LLM Process against conventional numerical methods on the S&P/TSX Composite — and why the quiet days flatter everyone while the loud days decide the ranking.

• Part 2 of 2

  Forecasting agents that read the news, and what it takes to trust them

  Hand the same task to agents that search and write a rationale. The hard part is leakage, the useful part is a handful of windows, and the honest ending is how little power a live market leaves you.

Around this project

• This repository on GitHub — the fork, including the experiment code under workshop_experiments/.
• The origin repo — the bootcamp foundation this project is built on.
• Architecture atlas — the origin’s map of the loop, the temporal fence, and the predictor families.