ADR-023 — Honest Positioning: Transparency/Discipline/Low-Cost Passive, Not an Alpha Machine

Date: 2026-06-28 Status: Accepted Related: 12-verification · 16-ta-analysts-experiment · 022-backtest-memory-contamination-and-holdout-discipline · 009-no-ta-style-analyst-agents

Context

After a systematic, pre-registered, cost-deducted, overfitting-resistant research cycle (signal-eval, PR #112), the question of whether the Committee/AI can generate tradable alpha has been empirically closed—the answer is no:

• Cross-sectional stock selection: Q1 univariate rank-IC (Holm-adjusted p=0.397) + M1 multivariate GBM (OOS IC +0.003, p=0.925), even on a survivorship-biased universe → The features read by the committee have no stock selection signal.
• Timing (Temporal): The Q2 finding that "returns are higher above the golden MA200" is beta—when turned into a tradable long/flat timing strategy (matching average exposure, deducting 0.38% costs, DSR deflated), it underperformed pure passive (terminal value 3.07 vs. buy-and-hold 15.10, Sharpe +0.36 vs. +0.68, drawdown −57% vs. −44%, even failing to "avoid downturns").
• Beyond Gold, Beyond Trends: Three assets (GC=F/510300.SS/NDQ.AX) × 4 signal families (trend/mean-reversion/volatility-target/breakout) × grid = 72 variants, none passed DSR>0.95 (the positive control cheating signal had DSR=1.00, proving the harness can identify true signals).
• News/Sentiment: ta-analysts (ADR-009) all three analysts performed below the base rate, "headlines are already priced in."
• Committee Closed Loop: Active trading underperformed passive + simple DCA (wiki 12 Negative #1).

The committee's directional prediction ability is negative: The latest automated audit (docs/verdict_accuracy.md) shows directional verdicts (excluding HOLD) have a true hit rate of 42.2% (n=56, below random). Its only marginally positive trait is default inaction: HOLD accounts for 56% → low turnover, less tinkering; don't mistake it for being able to "time the market/pick stocks" (its directional accuracy is worse than a coin flip).

Meanwhile, the README first screen (line 27) still uses "Cumulative Excess Return (Alpha) vs. 8 Benchmark Assets" as a selling point, while line 36 on the same page honestly discloses "audit tool, not a return-amplifying black box, directional hit rate 42.2% (below random), HOLD accounts for 56%." The first screen contradicts the evidence.

Decision

openInvest's positioning = Transparent / Disciplined / Low-Cost Passive Investment Companion, Not an Alpha Machine. Accordingly:

1. Do not claim or market alpha (including using backtested or live "excess returns" as a selling point). Implement the "public data red line / never market backtested alpha." The PnL area on the first screen retains transparent net value disclosure (showing performance itself is transparency, not a problem), but remove the "Excess Return (Alpha)" wording and selling point framework, replacing it with a neutral "Net Value vs. Benchmark Comparison (Transparent Disclosure, Not an Alpha Claim)", aligning the first screen with the honest disclosure on line 36. If OUTPERFORM_FEED is displayed, it must include both winning and losing trades (Red Line #3).
2. Anchor product value on three verifiable things:
   • Transparency/Auditability: telemetry (costs/latency/tool calls), committee transcript, decision_log, intervention counterfactual ledger—allowing users to see "why this decision was made."
   • Discipline (with cautious wording): Value lies in default inaction (HOLD 56% → low turnover) + defensive/sanity rules preventing impulsive actions; not directional accuracy (42.2% below random). Do not market "discipline" as "being able to predict direction."
   • Low-Cost Passive: Average DCA + core never-sell + low turnover, which was the unbeatable baseline in research → make it the honest default/recommendation.
3. Stop investing in "finding returns": No more committee stock selection tuning, more vol-target/sizing variants, or trend/regime timing tweaks marketed as alpha (all proven null). Scale-invariant per-asset regime thresholds (issue #113) are purely for robustness/self-hosting, not a return item.

Consequences

• README first screen alpha wording changed to neutral transparent disclosure (changed in this PR; outward-facing, final decision by owner via PR review).
• Subsequent product work shifts towards "transparency/discipline/low-cost passive": make the value of discipline (how many impulsive trades were prevented) and the audit trail visible to users; make average DCA + never-sell a clear recommended path.
• Research conclusions frozen in signal-eval + wiki 12; no more periodic token burning to re-evaluate "does the committee have alpha" (confirmatory theater).
• For fork/self-host users: Honest positioning reduces expectation mismatch of being treated as an "alpha black box," aligning with self-host distribution goals.

What We Will Not Do (Red Lines Continued)

• Do not market backtested/selected-window alpha; do not display specific hit rate numbers for n<30; public URLs must not contain fields that can reverse-engineer holdings (red lines unchanged).
• Do not interpret this ADR as "the committee is useless"—it has discipline + transparency value, it is simply not an alpha source, and positioning is calibrated accordingly.