EquityResearch

Description

EquityResearch is an environment for writing professional sell-side style equity research reports. Each task assigns the agent coverage of a real listed company (small/mid caps and investment trusts listed in the UK, Europe, North America and Australia). The agent researches the company from current public information using web search, builds its own forward estimates, derives its own valuation conclusion (an explicit fair value or target price with methodology shown), and submits a report (.docx or .pdf). The report is rendered to page images and text and graded per-criterion by an LLM judge against a two-tier rubric.

Each coverage brief and content rubric is reverse-engineered from a real published research note on the same company — which the agent never sees. Rubrics are deliberately time-robust: they check that the report covers the right company-specific substance and gets durable facts right, not that it reproduces any stale dated figure, so tasks remain valid as the companies' situations evolve.

Capabilities

• Web research on a live company (filings, results announcements, investor materials)
• Financial analysis: historical review plus the agent's own forward estimates
• Valuation: at least two approaches with assumptions shown, concluding in an explicit fair value / target price
• Professional document authoring (Word/PDF) with tables of estimates
• Analyst-style writing: conclusion-first, numbers-forward, catalysts with timing
• Long-horizon multi-turn execution (research → model → write → submit)

Compute Requirements

Agents are given a sandbox with 2GB of RAM and 2 CPUs, with network access for web research.

License

MIT

Tasks

Each task is one company coverage assignment. By default the agent receives a terse brief — the 2–4 sentence ask a real analyst would type (company, ticker/exchange, kind of note, "your own estimates and an explicit fair value/target price", page range) — so the model must supply the professional sell-side structure itself; the rubric grades it either way. A detailed brief variant (full section-by-section spec) is stored per task and can be selected via the task spec ("brief": "detailed"). Splits are assigned by company, so no company appears in both train and test.

• train: 269 tasks
• test: 31 tasks

Each task covers a distinct company (300 unique companies; one substantive reference note per company, median 11 pages).

Reward Structure

The submitted report is graded per-criterion by a gpt-5-mini judge (each criterion is one independent binary yes/no call over the report's page images and extracted text). Five rubric classes:

Structure covers the standard sell-side skeleton: header block, valuation conclusion with implied upside on page 1, a thesis stating the mispricing and catalyst, business description, industry positioning, financial analysis with estimates, ≥2 valuation approaches with assumptions, thesis-specific risks, ESG, and a programmatic 3–12 page check. Style covers conclusion-first writing, numbers-forward prose, active analyst voice, dated catalysts, estimate tables, concision and sourcing. Completeness/Correctness check company-specific substance and durable facts, judged with the real published note as reference context. Pitfalls are penalties for fabricated/unsourced material figures, a target price with no shown methodology, both-sides hedging with no investable view, multiple re-rating calls with no catalyst, format substituting for analysis, and stale data presented as current.

Reward = clip((points − penalties) / 100, 0, 1), with per-class credit proportional to criteria passed ("not applicable" verdicts are excluded from denominators). The agent is graded on the quality and internal consistency of the analysis, not on whether its call proves directionally right.

Calibration (golden tests): a real published professional note on the covered company scores ~0.79 with zero penalties; a hollow report with fabricated precise figures and an unsupported target scores 0.0 (its positive points are wiped out by pitfall penalties). On an end-to-end run of the same task, a strong general-purpose agent scored 0.87 with the detailed brief and 0.64 with the default terse brief — the gap is the internalized-format skill the environment trains.

Data

Coverage briefs and content rubrics are generated (gpt-5.5) from full-length research notes published free online; only substantive notes (≥8 pages) are used, at most one note per company. The note text is stored per-task as a judge-side reference only — it is never shown to the agent.

Tools

Agents get CLI tools (bash, read/write/edit, glob/grep, todo), Word, Excel and PDF toolsets for authoring, web_search and fetch_url (Tavily) for research, and submit_report for the single graded submission.

Time Horizon

Open-ended research-and-write tasks: the agent must research the company, build estimates, author a multi-page document and submit once. A capable agent typically needs on the order of tens of tool calls per task.

Environment Difficulty

The rubric is demanding: full credit requires company-specific completeness across ~10–14 generated criteria, factual consistency on ~8–12 more, the full sell-side structural skeleton, and professional style — while avoiding penalty criteria that specifically target the common LLM failure modes in financial writing.

Measured rewards with the default terse brief (3-task sample): gpt-5-mini mean 0.59, gpt-5.5 mean 0.70; gpt-5.2 scored 0.64 on a task where the ordering across all three models was strictly monotone in capability (0.55 / 0.64 / 0.72 vs the real published note's 0.79). Content Completeness (research depth) separates models most; Analyst Writing Style is uniformly the weakest class (3–5 of 8 criteria), leaving clear trainable headroom.

Other Environment Requirements

Two external secrets are required: openai_api_key (rubric judge) and tavily_api_key (agent web search).

Safety

Agents write research-style documents about real listed companies from public information. Outputs are training artifacts, not investment advice: there is no "right answer" target price, rewards do not depend on the direction of the call, and nothing is published or transacted. The reward design actively penalizes fabricated figures, unsourced claims and unsupported valuations, teaching agents to ground financial claims in verifiable sources. Agents have open web access for research; the sandbox is isolated and the only graded output is the submitted document.

Citations

Rubric and evaluation design draw on the CFA Institute Research Challenge written-report evaluation criteria, HealthBench (Arora et al., 2025) for signed-weight rubric scoring, ProfBench (Wang et al., 2025) for professional rubric schemas, and Rubrics as Rewards (Gunjal et al., 2025); the grading pipeline follows the PresentBench-style per-criterion multimodal judge.

@dataset{GREquityResearch,
  author    = {General Reasoning Inc. Team},
  title     = {EquityResearch},
  year      = {2026},
  publisher = {OpenReward},
  url       = {https://openreward.ai/GeneralReasoning/EquityResearch}
}