equity-investment-memo
equity-investment-memo
Description
Equity research benchmark — produce a one-page US-stock investment memo from a fixed local data snapshot, with mandatory chapter structure (Decision Card, Dual-Horizon Framing, Verified Facts, Derived Metrics, three-tier scenarios, Triggers, Judgment), arguing from both prospective-buyer and existing-holder viewpoints. Use when evaluating an agent's structured financial-analysis output. Originally authored as a SkillsBench task by Bingran You; copied here as a discoverable, self-contained reference.
SKILL.md
equity-investment-memo
A benchmark task for structured financial-analysis output. The agent is handed a snapshot of equity data on disk and must produce a tightly structured one-page investment memo — one that argues from two viewpoints, separates near-term timing from long-term ownership, and commits to a three-tier scenario stance with measurable triggers.
This is one of the SkillsBench tasks I wrote. The original — including
the data snapshot under /root/data/, the reference oracle, and the
verifier — lives at
BenchFlow-Hub/galaxies-bingran/tasks/equity-investment-memo.
What you'll find here is the instruction and the task config.
Instruction
You are an analyst, helping me for preparing an one-page investment decision memo for US stocks. You have a data snapshot for you to analysis. The snapshot is under
/root/data/folder.Only use data in
/root/data/. Do not do web search to get any external data. Today's date is indicated in the data set.Save final result as
/root/memo.md. It has to:
- Answer from the viewpoints of both prospective buyer and existing holder;
- Highlight recent (in the coming 1-2 quarters) and multi-year perspectives;
- Provide measurable triggers and three-tier scenarios: bull, base, and bear.
Strictly use the following chapter titles to organize the final markdown file:
## Decision Card ## Dual-Horizon Framing ### Near-Term Timing View ### Long-Term Ownership View ## Verified Facts (note: You need to quote the source URLs word by word in Verified Facts section.) ## Derived Metrics (note: for each line you need to display formulas or inputs ) ## Scenarios with ### Bull Case ### Base Case ### Bear Case ## Triggers ## Judgment
Task config
version = "1.0"
[metadata]
author_name = "Bingran You"
author_email = "bingran.you@berkeley.edu"
difficulty = "hard"
category = "finance"
tags = ["equity research", "investment memo", "structured output", "recent-ipo", "lock-up"]
[verifier]
timeout_sec = 600.0
[agent]
timeout_sec = 1200.0
[environment]
build_timeout_sec = 600.0
cpus = 1
memory_mb = 2048
storage_mb = 4096
Why it's a good test
- No web search allowed — forces the agent to reason from a fixed snapshot rather than wander the open web; rewards data discipline.
- Quote-the-source contract — the Verified Facts section must quote source URLs verbatim, which catches hallucinated citations.
- Two viewpoints, two horizons — the four-quadrant frame (buyer/holder × near-term/long-term) is a forcing function: a vague "this stock is fine" answer can't satisfy all four cells.
- Three-tier scenarios with triggers — the agent must commit to measurable conditions, not vibes.