Concluded hypotheses already live in concluded/ — keeping them in active.json
causes the registry to grow unboundedly. Runner now removes them at the end
of each cycle. Also cleaned up the existing social_dd concluded entry.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- iterate: CRITICAL note that all writes must use file tools, not narration
- iterate: new Step 5.5 — auto-registers one forward-test hypothesis per run
when a code change needs live-data validation
- research-strategy: cron changed from Monday-only to daily at 07:00 UTC
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Analysis of 25 picks reveals 60% 30d win rate (+2.32%) vs 41.7% 7d (-1.92%).
Score suppression is not the primary issue (avg score 71.5, 22/25 >= 65).
Root cause is evaluation horizon mismatch — ranker calibrated on 7d outcomes.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Statistical hypotheses now conclude immediately on the next runner cycle
without counting toward max_active. Only implementation hypotheses occupy
runner slots. Added conclude_statistical_hypothesis() for instant analysis
against existing performance data with Gemini LLM enrichment.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Implements ShortSqueezeScanner wrapping existing get_short_interest() in finviz_scraper.py.
Research finding: raw high SI predicts negative long-term returns (academic); edge is using
SI as a squeeze-risk flag when combined with earnings_calendar or options_flow catalysts.
Directly addresses earnings_calendar pending hypothesis (APLD 30.6% SI was strongest setup).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Implements ShortSqueezeScanner wrapping existing get_short_interest() in finviz_scraper.py.
Research finding: raw high SI predicts negative long-term returns (academic); edge is using
SI as a squeeze-risk flag when combined with earnings_calendar or options_flow catalysts.
Directly addresses earnings_calendar pending hypothesis (APLD 30.6% SI was strongest setup).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
gh pr create/list/edit were defaulting to TauricResearch (upstream).
Added --repo Aitous/TradingAgents to all gh calls in iterate,
research-strategy workflows and slash commands.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- active.json: updated days_elapsed from hypothesis runner
- hypotheses.py: black formatting applied by pre-commit hook
- .gitignore: local additions
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Uses google-genai SDK with gemini-2.5-flash-lite — same model already
used by the discovery pipeline, so no new secret needed (GOOGLE_API_KEY).
Removed ANTHROPIC_API_KEY from hypothesis-runner.yml.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- active.json: updated days_elapsed from hypothesis runner
- hypotheses.py: black formatting applied by pre-commit hook
- .gitignore: local additions
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Uses google-genai SDK with gemini-2.5-flash-lite — same model already
used by the discovery pipeline, so no new secret needed (GOOGLE_API_KEY).
Removed ANTHROPIC_API_KEY from hypothesis-runner.yml.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
days_elapsed counts entries in picks_log, so running on weekends would
inflate the counter with noise picks. Exit early on Saturday/Sunday.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Before concluding a hypothesis, check if the scanner's source file
changed on main since created_at. If it did, the baseline picks in
performance_database.json reflect the updated code for the later part
of the experiment, which can confound the comparison.
When drift is detected, a warning is embedded in:
- the concluded .md doc (blockquote below Decision)
- the PR comment (blockquote in the conclusion body)
The programmatic decision is not overridden — the warning is purely
informational, allowing the reviewer to judge whether the result is
trustworthy.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
load_dotenv() in tradingagents/config.py searches the cwd for .env.
Worktrees in /tmp/ don't have one, so symlink the main repo's .env
into the worktree root before running discovery.
In CI, secrets are passed as env vars directly — symlink is a no-op.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Pending hypotheses queue by priority and promote when a slot opens,
rather than pausing a running experiment mid-streak.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Youssef Aitousarrah
Aitous
github-actions[bot]