README.md

metadata

license: cc-by-4.0
pretty_name: Herculean post-release window
tags:
  - finance
  - agents
  - benchmark
  - mcp
  - trading
  - look-ahead

Herculean post-release window (Aug–Sep 2026)

Market environment and task instances for evaluating the Herculean market workflows on a window that starts after every evaluated backbone was released, so the outcomes cannot be in any backbone's weights.

• Window: 2026-08-03 to 2026-09-30 (42 trading days). Latest backbone release: Claude Sonnet 5 (2026-06-30).
• Workflows: Trading (4 stocks x 42 days), Hedging (8-stock universe, 42 days), Market Insights (4 stocks x 9 weeks).
• History available to agents: December 2024 onward; all tool queries are bounded to the decision date.

Files

file	content
env.duckdb	the environment database read by the MCP exam servers (prices, news, filings), Dec 2024 – Oct 9, 2026
data/prices.parquet	daily OHLC, close, adjusted close, volume
data/news.parquet	one highlights summary per symbol and calendar day
data/filings.parquet	10-Q / 10-K MD&A and Risk Factors sections
trading/, hedging/, insights/	instances.jsonl: one trajectory per case with its decision dates
post_selection.json	the configurations evaluated in this window and the rule used to select them

Rows up to 2026-03-31 are the original Herculean environment (TheFinAI/Herculean); rows from 2026-04-01 are new.

Sources for the new rows (2026-04-01 onward)

• Prices: Yahoo Finance daily bars via yfinance (unadjusted OHLC plus adjusted close).
• News: Google News RSS searched per symbol and day, keeping only articles published that day, summarised with the Agent Market Arena prompt (testbed/get_daily_news.py in The-FinAI/Agent_Market_Arena) on gpt-5-nano, the model that produced the original highlights. RSS provides headline, source and date only, so these summaries are built from headlines rather than article text.
• Filings: SEC EDGAR 10-Q / 10-K filed in the window; MD&A (10-Q Item 2, 10-K Item 7) and Risk Factors (Item 1A).

Configuration selection

The configurations run in this window are the top three (harness, backbone) pairs per workflow by mean annualised Sharpe over three runs in the offline window (2026-01-02 to 2026-03-31), fixed before any post-window run (post_selection.json). Run outputs are mirrored in TheFinAI/herculean-runs under post/.