OpenInvest

English

中文

English

中文

Appearance

ADR-009: Reject TradingAgents-style Analyst Subagent Extension ​

Date: 2026-06-11 Status: ✅ Accepted (based on test_ta experiment, pre-registered judgment)

Context ​

The core paradigm of TradingAgents (TauricResearch) is: multiple analysts (fundamental/sentiment/news/social media) produce reports → researcher debate → trader decision. When openInvest aligned its data dimensions on 2026-06-09, it chose "deterministic fact blocks into reasoning, no voting agents" (see wiki 02/15), but that choice was a design judgment at the time, not experimentally validated. The user requested a practical test: whether adding 3 TradingAgents-style analysts yields any benefit.

Decision ​

Do not introduce report-based/voting-based analyst subagents. Maintain the current architecture: deterministic fact blocks (regime probability caliber / valuation / sentiment dashboard / path reference) injection + LLM translation interpretation + probability table anchoring direction + parse_cio_memo deterministic post-processing.

Rationale (test_ta Phase A, full text in wiki 16) ​

Pre-registered Gate (30d hit rate Wilson 95% CI lower bound > max(subset directional base rate, input data mechanical mapping)):

  • fundamental: 50.3% ≈ coin flip (the gold side used real CFTC COT data—the data itself has no signal)
  • news: 57.4% vs base rate 73.0% (GDELT real headlines, title-only downgrade)
  • sentiment: 64.7%—beat its own input mechanical mapping (57.5%), but lost to the "blindly bullish" base rate of 73.1%

0/3 passed the line → integration stage (Phase B/C) not triggered per pre-registration. n=244/analyst, same 122 decision days as historical ablation, zero lookahead, ~¥10-20 cost.

Implications and Boundaries ​

  • LLM's incremental value lies in interpretation, not direction voting (sentiment beat mapping but lost to base rate proves this)—supports the "fact blocks + CIO veto" approach
  • Single bull market window conclusion; bear/sideways window not covered → Re-tested on 2026-06-12 (wiki 16 §5b): 2022 bear market window × 2 vendors × 3 models (MiMo/DeepSeek) + ensemble of three combinations. Fundamental passed the Gate in a single cell (2022×flash), but only 1/3 across models (reasoner/ds-chat both converged to mechanical mapping level)—according to the pre-registered "2/3 counts as signal" standard, a single piece of evidence is insufficient, original ruling stands with stronger evidence
  • Conditions to overturn this ADR (tightened after re-test): an analyst passes the pre-registered Gate and is reproducible across ≥2 vendor models → run Phase B integration A/B (code already prepared in test_ta branch, flag off by default). A single model passing has been shown to be potentially a coincidence of window × model style (flash-2022 precedent)
  • Residual clue: ③ mechanical mapping (COT 4-week change + valuation percentile) in 2022: 61.3% vs base rate 51.4% (n=62, not significant, but the convergence point for all three models)—if validated later, implement as a deterministic fact block, not an LLM agent, and add to the forward experiment queue
  • News dimension backtest fully closed (2022 two models ≈ coin flip, headlines already priced in); fair test still pending in the forward experiment queue (accumulating event-layer data)
  • wiki 16-ta-analysts-experiment.md (full experiment text)
  • wiki 12-verification.md (core claim empirical checklist)
  • Existing conclusion "regime is signal, verdict is noise" (verdict direction ≈57% blind) mutually corroborates with this experiment