⚠ TRUTH LAYERDashboard vs DB-reality gap·Per-class tiles below show post-resolver-v2 filtered performance.RAW at_raw_picks (Kimi + 2026-05-12 audit): 11.13% WR / PF 0.46. Refreshed 2026-05-29 (n=73,817 closed): WR 25.28% / PF 0.61 (outlier-capped) — uncapped PF 2.5 is a handful-of-rows outlier artifact, do NOT cite it. Either read: RAW is net-unprofitable.
No asset class currently passes industry-standard real-money thresholds (DSR>0.95, PBO<0.05, WFE>60%, live-Sharpe>0.5, n≥100). The earlier DSR=1.0 candidate (cot_positioning + CT=F) was FALSIFIED on 2026-05-13 after dedup audit found 6.33× over-emission (38 raw signals from only 6 unique CFTC release weeks); pilot status is now SHADOW_INSUFFICIENT_N with dsr=null. See paper_pilot.html for live pilot JSON and reports/cot_paper_pilot_overemission_falsified_20260513.md for the autopsy. real_money.html still gates the 10-step Lopez de Prado readiness; NFA — no real-money sizing without explicit greenlight + all gates clear.
✅ DATA QUALITY IMPROVEMENTS (2026-05-31):
Ghost rows: 0 (was 22,947). Status-PnL mismatches: 0. Non-canonical statuses (WON/CLOSED_SL/CLOSED_TP/FLAT): all 0. Backup tables: removed.
at_raw_picks UNKNOWN: 3 (was thousands). FOREX: PR #6 policy-gate shipped (all blocked except cta_cross_asset_tsmom SHORT).
BOND: now emitting (78 raw picks). See updates/2026-05-31-pr1-data-integrity-repair.md.
⚠ COMMODITY filter-survival gap (2026-05-17):
dashboard shows WR=85.5% (n=228) vs raw WR=60.2% (n=354). The −126 difference is picks excluded by
_is_valid_resolved_pick() (missing / corrupt pnl_pct — concentrated in losing cta_replicator
and combined_confidence trades). The 85.5% reflects only picks with validated pnl data, not the full cohort.
Raw 60.2% is the conservative estimate. FOREX: WR=57.8% but PF=0.85 — wins avg 0.62% vs losses avg 1.00%
(tight-TP / wide-SL asymmetry); total PnL=−15.84% confirms it is net-negative despite majority wins.
Score-booster now applies per-class penalty/reward at alpha_engine/score_booster._calibrate_confidence()
(CRYPTO >0.85 = −12, <0.6 = +3; EQUITY 0.85-0.90 = −15; FOREX ≥0.85 = −10).
Until upstream generators recalibrate, do not size off raw confidence — weight off post-booster score.
🎯 MAJOR GOALPhenomenal performance across ALL asset classes — sustainable, hedge-fund-grade.
All 4 inflated CRYPTO nav surfaces (PF 9-18, WR 78-88%) are dominated by historical claude_gainer_st picks (91.7% concentration) with EXPIRED→WON resolver mislabels. Independent DB probe shows raw equivalent WR=33%. Honest cells: pf_registry policy-clean-net PF 0.98 / WR 39.7% / n=229; money_ready_verdict PF 0.96 / WR 31.3% / DSR 0.0073. See pick_funnel.html DISPUTED banner and reports/ASSET_CLASS_EDGE_FIX_PLAN_2026-05-27.md. Do not size on the inflated surfaces.
Tier definitions: T1 PF>2/WR>55/MDD<10 (Renaissance). T2 PF>1.5/WR>50/MDD<20 (Institutional). T3 PF>1.2/WR>48/MDD<30 (Retail-OK). Source: asset_class_health in audit_dashboard/data/dashboard_data.json (post-resolver-v2 noise filter, generated 2026-05-05T01:37Z). Resolver v2 shipped 2026-04-28; v2.1 bug bundle 2026-05-02. Historical re-resolve (re_resolve_historical_v2.py) pending — pre-fix labels still in closed_picks for picks closed before 2026-04-28.
⚠️ Two PF/WR figures may appear per class on this page. The card above pulls asset_class_health (full-history closed-trade aggregate, post-resolver-v2 noise filter). The lower-down hf_stats.by_asset_class panel uses a recent-subset window (typically last 60-90d closed picks). Recent figures often diverge from headline — CRYPTO recent PF 0.89 vs headline 1.25 (n=1650 vs n=8067), COMMODITY recent 1.09 vs 1.78. For deploy/sizing decisions, weigh the recent panel. For long-horizon strategy validation, weigh the headline. Source: dashboard_data.json::hf_stats.by_asset_class.
📊 Per-asset-class walk-forward (OOS)Out-of-sample metrics from walk_forward_by_class() — Sharpe colored: green > 0.5, yellow 0–0.5, red < 0.
Loading…
📊 Strategy IC AnalysisSpearman rank correlation between elite_score and PnL outcome — does the scoring model actually rank winners above losers?
Loading…
🏆 TIER-2 PROVENBuried high-edge strategies promoted from the alphabetical systems grid. Tier badges per CHARTER §2.
📊 TA BASELINETradingView 6-strategy benchmark (ema_cross/macd/rsi/bollinger/donchian/supertrend) vs buy-hold. Promotion gate: walk-forward robustness ≥0.60 AND trades ≥5.
PROVEN — A pick qualifies as PROVEN when its strategy has: (1) ≥5 closed trades, (2) ≥55% win rate after Bayesian shrink, (3) Profit Factor ≥1.5, and (4) confidence ≥0.7. Manually vetted systems (e.g., alpha_engine, battleground+) can also earn PROVEN status. Smart Picks — An AI-curated basket. Every active pick is scored on 6 dimensions: Regime match (25%), Quality score (35%), Freshness (15%), TP upside remaining (15%), Higher-timeframe alignment (10%), plus a Proven Winner Boost (+8–15 pts). Only the top-ranked picks make it in. Verified Alpha — Signals that are auditable and verifiable: prediction-market consensus (Polymarket/Kalshi), audited copy-trader clones, or any strategy with ≥55% forward win rate on ≥5 trades. High Conviction — The strictest preset. It applies shared hard gates (score ≥40, trust tier, forward WR, regime alignment, consensus) plus per-asset-class floors. If a row shows here, it passed every single gate. Trust Tiers — PROVEN = elite edge. DEVELOPING = good edge. WATCH = marginal. SANDBOX = unproven/new. PROBATION = broken/banned system.
🏆 SUPREME EDGE 2026-05-12 refresh — DSR-verified real edge per asset class
⚠️ RESEARCH ONLY — NOT ACTIONABLE FORWARD SIGNALS. The DSR numbers below are derived from closed-pick historical analysis. Post-hoc segment search, over-emission artifacts, and small-sample DSR inflation (n<100) affect multiple entries. Do NOT use for live sizing without independent forward-test confirmation. See INC #1 (P0) post-hoc segment warning below.
Latest patterns leading to strong profit and low risk, surfaced via the new Anti-Overfit Audit (DSR) sidecar (Lopez de Prado AFML eq 14.5; hourly cron). 8 of 42 strategies (n≥20 closed) pass DSR ≥ 0.95 floor — "publishable confidence" per Bailey-Lopez de Prado.
COMMODITY — cot_positioning DSR=1.0 FALSIFIED 2026-05-13: the original "n=104 / WR 86.5% / DSR=1.0000" claim was an over-emission artifact — dedup audit found multi_asset_cot_positioning emitted 38 raw signals from only 6 unique CFTC release weeks (6.33× ratio). Pilot status now SHADOW_INSUFFICIENT_N with dsr=null; dedup pilot WR 33.3% on n=6 unique releases. Do not size on this edge. Refs: reports/cot_paper_pilot_overemission_falsified_20260513.md, audit_dashboard/data/cot_paper_pilot_status.json. CRYPTO — ML-enhanced 1d/1h sleeves: 4 strategies survive DSR test — ml_enhanced_INJUSDT_1d_B_lightgbm (n=27, WR 100%, Sharpe +2.49), ml_enhanced_DYDXUSDT_15m_D_ensemble_stack (n=31, WR 96.8%), ml_enhanced_FETUSDT_1d_B_lightgbm (n=25, WR 100%), ml_enhanced_RENDERUSDT_1h_D_ensemble_stack (n=34, WR 85.3%). All DSR ≥ 0.9995. ⚠ All four have n < 100 — DSR formula is unreliable below n=100 (Bailey-Lopez de Prado). Awaiting n≥100 confirmation before sizing. EQUITY — stocks_rsi2_pullback DISPUTED (INCIDENT #96, 2026-06-04): original "n=70 WR 62.9% +0.78%" claim was raw-row pre-dedup. True n=140 (deduped by sym+ts+strat+dir) WR 33.9% avg -0.007% — slightly losing, not edge. Per-symbol: WMT n=220 avg -0.44%, KO -0.80%, COST -0.59%; only C/AVGO marginally positive. Inflation source: INCIDENT #91 NULL opened_at row duplicates (5.6x ratio). 15-minute timeframe is overfit-bait: 33 of 42 audited strategies are OVERFIT_LIKELY (DSR < 0.5). Pattern: ml_enhanced_*_15m_* family fails the multiple-testing correction even when raw Sharpe looks good. Avoid any 15m-tagged ML strategy as a standalone edge claim. FOREX has ZERO DSR-survivors: post-#876 unit clamp + #891 closed_at repair, no FOREX strategy passes DSR ≥ 0.5. Master plan FOREX state = BLOCKED. Elite-score floor raised 50→70 (commit 4a2d337a5dc) acts as admission hard-cap. Do not size FOREX even if forward WR looks recoverable — the multiple-testing bar is unmet. ML calibration is SYSTEM-WIDE INVERTED (Kimi finding verified P0 #9 via direct DB probe): conf≥0.9 bucket WR 14.4%, conf 0.5-0.6 bucket WR 60.3%. Confidence is currently an anti-signal across asset classes, not just ETF/CRYPTO. Mitigation: _normalize_confidence helper shipped at audit_trail/quality_gates.py defends against the additional schema bug (~2k rows on 0-10 scale leaking into 0-1 readers). Real-money posture as of 2026-05-31: COMMODITY/EQUITY are REHAB (STABLE_EDGE per edge-stability sidecar); CRYPTO has DSR-verified ML sleeves (see anti-overfit audit); FOREX is BLOCKED (PR #6 hard-gate, PF 0.29 / WR 46.1%); FUTURES remain BLOCKED; ETF/BOND/INDEX_STOCK are sample-size-thin. NFA.
🎯 Where Our Edge Actually Is (Closed-Pick Data, n=4,618)
⚠️ POST-HOC SEGMENT WARNING — the numbers below (82% WR / PF 11+ etc.) were surfaced by scanning ALL closed-pick confidence-band slices for the best-looking cell, after the fact. This is post-hoc segment-search and IS NOT an actionable forward signal. Multiple-testing correction (Bonferroni / Benjamini-Hochberg) is not applied; the implicit search-space >50 cells means an 82% WR cell can arise from pure noise on a true null. Treat these as historical pattern descriptors, not predictive filters. The DSR-verified strategies above (Bailey-Lopez de Prado AFML eq 14.5) are the methodology-clean version. INC #1 (P0) — opencode/ring-2.6-1t + 3-engine swarm, codified 2026-05-26.
Crypto Confidence 0.85–0.90 — (post-hoc cell)82% WR, PF 11.8 (n from per-class audit). Crypto >0.90 hits an overfit cliff (47% WR). Proven ML Strategies — 8 ML-enhanced strategies with n≥5, WR>55%, PF≥1.5: 79.4% WR, avg +0.08%, PF 11.34 (n=199). Top: DYDX 15m (95.5% WR), STRK 15m (95.2% WR), INJ 1d (95.0% WR), BNB 15m (89.5% WR). Proven + High Confidence Combo — PROVEN strategy + CONF 0.8–0.9: 71.3% WR, avg +0.11%, PF 13.21 (n=94). R:R Truth (CRYPTO, verified 2026-04-17 across 1,916 closed picks) — R:R 1.0–1.5: 62.3% WR, PF 1.66, +0.71% avg (n=150 — highest WR band). R:R 1.5–2.0: 52.5% WR, PF 1.92, +0.69% avg (n=983 — volume sweet spot). R:R ≥2.0: 58.0% WR, PF 3.06, +0.99% avg (n=715 — highest PF AND avg). R:R <1.0: 55.9% WR but PF 0.93, -0.19% avg (n=68 — high WR can't overcome bad geometry). Prior tooltip claimed R:R≥2.0 was 29.5% WR catastrophic — empirically wrong; triple-verified via DeepSeek + Inception mercury-2 + Ollama Cloud gpt-oss:20b. Confidence is NOT global — FOREX peak is 0.75–0.80 only (49% WR); 0.70–0.75 is DANGER (25% WR). EQUITY is bipolar: >0.90 works (67% WR) but 0.85–0.90 is the WORST bucket (20% WR). COMMODITY peak is 0.70–0.75 (48% WR). A single confidence threshold cannot work across asset classes. Direction = BUY — 3,909 picks, 28.9% WR, PF 0.38. Direction = LONG — 441 picks, 54.9% WR, PF 3.14. (The winning cohort is signal_type=BUY + direction=LONG at 62.6% WR.) High-grade A/B — NOT an edge: 49.3% WR, PF 0.66, -0.08% avg (n=483). Asset Class Edge — ML-Enhanced (crypto) = 55.1% WR, PF 1.77. Quan Engine (crypto) = 29.0% WR, PF 0.38. Non-crypto closed-book data is thin or unprofitable; size smaller or avoid. Bottom line: For crypto, prioritize ML-enhanced proven strategies and confidence 0.85–0.90. Avoid quan_engine scalps, wide R:R, and BUY-only signals. For non-crypto, be aware that tile PnL is now cost-adjusted (net), which can flip the sign on thin edges.
Best Filter by Asset Class
Asset
Use This Filter
WR (filtered)
WR (all picks)
Lift
PF (filtered)
Crypto
Confidence ≥0.8 or Proven strategy
69.2%
49.6%
+19.6pp
10.39
Stocks
Trusted + score ≥50
69.2%
43.4%
+25.8pp
0.77*
Forex
Trusted
49.0%
42.9%
+6.0pp
3.59
Commodities
Trusted
44.5%
42.3%
+2.2pp
1.26†
* Stocks Trusted produces a high WR but PF 0.77 on a small n=13 — one outsized loser offsets many small wins. Not a verified edge by the PF > 1.5 rule; combine with Maximum Conviction Combo (PROVEN + score ≥ 50), which matches the closed-book equity filter in MERCURYPROMPT.md (Score ≥ 50 + Trust ≥ 3, PF 2.62). † Commodities Trusted: PF 1.28 with bootstrap CI straddling 1.0 on n=273 — inconclusive in MERCURYPROMPT.md; listed here for WR lift only, not gold-standard edge. R:R ≥ 1.5 filter currently underperforms baseline across every asset class (crypto −0.4pp, equity −1.0pp, forex −9.2pp, commodity −32.3pp). Use cautiously.
Maximum Conviction Combo (under re-validation)
PROVEN strategy + confidence 0.8–0.9 — insufficient sample Original claim (71.3% WR, PF 13.21, n=94) is not reproducible on the current recent_closed window (n=0 matching picks). The band is hidden from display until it repopulates ≥20 closed trades under the post-correction PROVEN tier (claude_gainer_st was demoted 2026-04-20). Re-evaluate weekly. See docs/AUDIT_EFFECTIVENESS_AUDIT_2026_04_20.md and docs/REMAINING_ENHANCEMENT_PROPOSALS_V3_2026_04_20.md for the methodology under reconstruction.
What the Badges Mean
🏆Forward-proven — strategy verified in BOTH backtest AND forward testing. The real deal. ⚠️BT-forward gap — backtest looks great but forward performance is weaker. Proceed with caution. (n=X)on Profit Factor — small sample warning. PF on <30 trades can be dominated by one lucky trade.
What to Avoid
•Grade D & F picks — 725 trades at 33.4% WR, PF 0.82, −131.8% cum PnL (improved from −509% after blocked-source sweep) •Commodities SHORT on copy_trader — 50% of picks come back as LOST (unresolved), masking real TP/SL performance •Futures — 6.3% WR on n=17 with 76% LOST-exit rate. Kill list pending. •ETFs — PF 0.28 on 19 trades, dominated by extreme_oversold_bounce at 0% WR •Low-confidence crypto — confidence <0.6 shows 26–44% WR and deep negative avg PnL •R:R ≥ 1.5 filter on non-crypto — currently harmful across equity, forex, commodity (see footnote above)
Understanding Profit Factor
PF = gross wins ÷ gross losses. PF > 1.5 on adequate recent_closed sample with PF CI above 1.0 is our verified edge bar (same criterion as MERCURYPROMPT.md). PF > 2.0 = strong. PF > 3.0 = exceptional. HIGH CONVICTION on Active Picks enforces forward (FWD WR) + score + trust gates from that evidence so the live table stays consistent with these PF-verified cohorts (crypto / equity / forex only in strict mode).
The Avg PnL/trade row grounds the Realized PnL sum — a +621% PnL on 1,745 trades is really +0.36% per trade on average.
How the feeds stack up: ●Verified Alpha — PROVEN trust tier from vetted source systems. Bimodal: sub-cohort quality varies; check per-strategy stats before sizing. ●Smart Picks — passes the strictest per-asset gates (min score, RR, forward WR, regime alignment). Tightest filter currently live. ●High Conviction — forward-validated tier: FWD WR ≥55% (≥70% crypto/equity/forex), score ≥ asset floor, ≥5 forward trades, trust tier clean. Applies hc_filter.js gates + per-asset S/A/B contract. ●Active Picks — all live picks that pass hygiene (dead-symbol guard, blocklist, status checks). Broadest feed. Risk-adjusted metrics (Sharpe, max DD, net-of-cost PF, regime decomposition) pending — no headline point-estimate claims until they land. See docs/REMAINING_ENHANCEMENT_PROPOSALS_V3_2026_04_20.md.
Strategy History
🔥 HIGH CONVICTION FILTERS APPLIED
Active Picks is now filtered to picks that pass both the shared hc_filter.js gates (score≥40 compound, trust, forward WR, regime, consensus) and the per-asset-class gate below. Those gates are tuned so live rows match cohorts that showed verified closed-book edge (Profit Factor > 1.5 with bootstrap CI above 1.0 on recent_closed — see MERCURYPROMPT.md and the ? Guide overlay for the same definition).
Not a second kind of “edge” — HIGH CONVICTION does not invent a different standard. It applies forward-testable thresholds (FWD WR, Score, Trust) on each active row so only asset classes with PF-verified history (crypto, equity, forex) appear in strict mode; weak / dead classes are hidden.
Field reference — the thresholds below map 1:1 to columns in the Active Picks table so you can filter/sort manually even when no pick fully passes:
• Score = the SCORE column in Active Picks (field pick.score, Dashboard Score 0-100).
• Trust = the TRUST column (field pick.trust_score, 0-10).
• Forward WR = the FWD WR column (field pick.strat_fwd_wr, %). N = FWD N column (strat_fwd_trades).
If any of these columns aren't visible, open the ⚙ column settings button on the toolbar to enable them.
Closed-book reference table: MERCURYPROMPT.md § Current Edge (2026-04-14, ~3,500-pick recent_closed denominator) — PF, WR, and bootstrap CI. Per-class lines above summarize forward WR lifts for the same cohorts; the gold-standard check remains PF > 1.5 with CI clear of 1.0.
Classes marked DEAD, WEAK (commodity: inconclusive PF), or NO DATA are excluded from strict HIGH CONVICTION — click ✕ to return to the full view.
Want near-misses? Close this panel, sort Active Picks by Score descending, then filter FWD WR ≥ 45%.
Claude Top Picks Curated by Claude Opus 4.6
3 highest-conviction picks selected via multi-factor analysis: RSI positioning, volume liquidity, momentum quality, Genesis score, and sector narrative. Updated each scan cycle.
Entry/TP/SL levels auto-tracked against live Binance prices. Historical performance builds over time as rounds complete.
Loading...
5-AI Top Picks Battle Claude vs Antigravity vs Grok vs KIMI vs Mercury
Live paper trading competition. 5 AI agents selected their top picks independently. Prices auto-refresh from Binance. Who will win?
Loading...
Score Tracker - What-If Performance
Every 15 minutes, the dashboard snapshots the top-scored picks. This tracker shows how those picks performed over time - answering "what if you traded the top picks by score?"
Snapshot History (newest first)
🧠 Smart Picks
?Loading...
💡 In Simple Terms
Smart Picks are our highest-conviction trade ideas. Instead of showing you every signal, our AI analyzes thousands of data points — past performance, market conditions, strategy track records — and only surfaces the picks that have historically led to profitable trades. Think of it as a filter that learns from every trade we’ve ever made to find patterns that actually make money.
⚙️ Technical Details
Each pick passes through a multi-layer scoring pipeline trained on 3,500+ closed trades:
█ML Score (0-25 pts) — XGBoost model (IC=+0.33) trained on 40+ features including volatility regime, orderbook imbalance, and cross-asset correlations
█Forward Walk-Forward WR (0-40 pts) — Out-of-sample win rate from rolling 30-day windows. Strategies must prove they win on unseen data, not just backtests
█Confidence Calibration (0-20 pts) — Sweet spot is 0.60-0.85. Over-confident signals (>0.90) are penalized — our data shows they have 19.5% WR
█Regime Alignment — Is the pick’s direction consistent with the current macro regime (bull/bear/neutral)?
█Trust & Source Scoring — Each source system and strategy has a data-driven trust score based on realized P&L across hundreds of trades
Hard gates then filter out: SCALP mode (24.8% WR), panic-health entries, strategies with <35% forward WR, direction conflicts, and 80+ permanently killed strategies identified through continuous monitoring.
Forward-validated bypass: Strategies with ≥20 trades and ≥50% forward WR can qualify even if their composite score is below the threshold — real results outweigh theoretical scoring.
ℹ How Smart Picks Work: Every scan cycle, our engine scores all active picks on 6 dimensions:
Direction matches regime (25%), Quality score (35%), Freshness (15%), TP upside remaining (15%), Higher-timeframe alignment (10%), Proven Winner Boost (+8-15pts).
NEW GATES: Crypto shorts blocked (15.3% WR live), Confidence floor at 0.70 (below = 23% WR), 14 banned systems auto-excluded.
NEW SOURCE: Gainer Momentum Interceptor — intercepts top 24h gainers with confirmed momentum continuation.
? Glossary
Scoring Factor Glossary — what each item in the "Why" column means:
Quality score 62.0 (+22)
The elite scorer rated this pick 62 out of 100 based on win rate, Sharpe, forward confirmation, and trade count depth. Higher = more reliable strategy track record.
Fresh signal 3.4h ago (+12)
The signal was generated 3.4 hours ago. Signals under 4h old get a freshness bonus because stale picks decay in edge.
92% upside to TP (+15)
The current price is 92% below the take-profit target. More room to TP = stronger upside potential and better risk/reward entry.
Higher-timeframe confirms LONG (+10)
The 4h or daily timeframe trend agrees with this pick's direction. When the higher timeframe opposes, the pick gets 0 points instead.
Multi-timeframe STRONG 3/3 (+10)
3 out of 3 timeframes (1h, 4h, daily) agree on direction. STRONG = 3/3, MODERATE = 2/3, WEAK = 1/3 or less.
Multi-signal agreement (+20)
Multiple independent strategies or indicator families agree on this pick. Cross-family confluence strengthens conviction.
Proven winner (+15)
This strategy is on the proven-winners list (manually vetted strategies with strong, statistically significant forward track records across many symbols).
Copy-trader premium (+N)
This pick aligns with positions taken by tracked profitable copy traders on Bitget.
Institutional sweet spot (+10)
Position size falls in the range where both retail and institutional traders can enter without excessive slippage.
Direction matches BULL regime (+N)
The pick's direction (LONG/SHORT) aligns with the current market regime. Fighting the regime incurs a penalty.
Regime penalty NEUTRAL (-8)
The current regime (BULL/BEAR/NEUTRAL) doesn't support this pick. NEUTRAL regime penalizes all picks; BEAR penalizes LONGs.
Short boost: fear & greed=11 (+15)
Contrarian SHORT bonus when the Fear & Greed Index is in extreme fear (<20). Crowded longs get squeezed, shorts have edge.
Factor + post-earnings drift (+5, BULL)
Equity factor model detects a favorable fundamental factor (e.g. value, momentum) combined with post-earnings announcement drift in the pick's favor.
ASSET CLASS:
🎯 What Makes a Winner - Decile Test Results
1,927 closed picks
▶ click to expand
Top 20% ml_score
60% WR
+4.2% avg PnL
vs
Bottom 20% ml_score (<0.50)
22% WR
-8.14% avg PnL
▶ Predictor Power Ranking (Spearman rho = rank correlation with win/loss outcome)
Rank
Scoreⓘ
Spearman rho
Bottom 30% WR
Top 30% WR
Spread
Verdict
🥇 1
ml_score
+0.33
32.5%
60.0%
27.5pp
BEST PREDICTOR
🥈 2
confidence
+0.27
30.8%
51.9%
21.1pp
STRONG
🥉 3
elite_score
+0.012
28.0%
39.4%
11.4pp
WEAK
⚠ Overconfidence Pattern: D10 (top decile) underperforms D9 in both confidence and ml_score.
The highest-scoring picks are slightly worse than the second-highest tier. Extreme scores signal overconfidence, not edge.
ZEROED ✘Backward-looking momentum; double-counted with smart_picks
Monte Carlo / Meta-Label / Hindsight / Skyrocket
various
n/a
ZEROED ✘Cross-AI consensus: survivorship bias / unvalidated / broken ML
Raw score normalized to 0-100 via round(raw * 100 / 90).
Overconfidence cap: score >85 with <10 closed trades is capped at 60.
Fallback: when full scorer collapses (sparse data), a 3-factor model (60% strategy WR + 30% regime + 10% R:R) takes over.
Hard Quality Gates (must pass ALL to enter Smart Picks)
R:R ≥ 1.2Unfavorable risk/reward filtered
TP remaining ≥ 10%Near-target picks excluded
Age ≤ 48h(168h for copy traders)
TP/SL not hitAlready-resolved picks excluded
No meme LONG in bearDOGE/SHIB etc. long blocked in bear regime
Banned systems excludedvol_spike_backfill, winner_pattern, etc.
Soft Filters (score-dependent, may block weak picks)
Wrong direction + score <70Counter-trend with low score = filtered
MTF BLOCKED + score <55Multi-timeframe contradiction on weak picks
Ensemble <2/3 + score <65Independent signals disagree on mediocre picks
Non-crypto quarantineMax 3 forex/equity picks; must be PROVEN trust tier
📈US Equity Picks (UEPS)
Building track record · n=0/100
Long-term value (1y–3y+) and swing (1w–1m) US equity picks. Composite scoring per findings/SYNTHESIS.md §3:
0.55 × ValueComposite + 0.45 × QualityComposite × SafetyGate
where ValueComposite blends Magic Formula + Acquirer's Multiple + FCF Yield, and SafetyGate requires Altman Z'' ≥ 1.10 AND Beneish M ≤ -1.78.
Tier classification deferred until n≥100 closed picks per docs/PERFORMANCE_CHARTER.md §10.
Resolver (long-term) never closes on price drawdown alone — thesis-break + IV-attainment + horizon expiry only.
UEPS Fundamentals Glossary
F-Score (Piotroski) /9
Joseph Piotroski's 9-point fundamental quality score (2000). Adds 1 point each for: positive ROA, positive operating cash flow, ROA improvement, OCF>net income, lower long-term debt ratio, higher current ratio, no new shares, higher gross margin, higher asset turnover. Score ≥7 historically outperforms; ≤3 underperforms.
Magic Rank (#)
Joel Greenblatt's "Magic Formula" rank (2006): combined ranking of earnings yield (EBIT/EV) and return on capital (EBIT/(NWC+net fixed assets)). Lower number = better combined rank. Top-30 typically forms the Magic Formula portfolio.
Acquirer M (Acquirer's Multiple)
Tobias Carlisle's enterprise multiple: Enterprise Value / Operating Earnings (EBIT). Lower = cheaper from an acquirer's perspective. Carlisle's research: lowest-decile Acquirer's Multiple beats the Magic Formula on 26-year US backtest. Typically <7x is "deep value".
Altman Z'' (safety)
Edward Altman's Z-score variant for non-manufacturers: 6.56(WC/TA) + 3.26(RE/TA) + 6.72(EBIT/TA) + 1.05(BV equity / total liabilities). Z'' ≥ 2.6 safe; 1.10-2.6 grey zone; <1.10 distress. SafetyGate requires ≥ 1.10.
Return on Invested Capital = NOPAT / (Debt + Equity - Cash). Measures how efficiently a business converts capital into after-tax profit. Sustainable ROIC > cost of capital (typically >10-12%) is the hallmark of a compounder.
FCF Yield
Free Cash Flow / Enterprise Value. The cash-on-cash yield an acquirer would receive at today's price. >6% historically rich; 3-6% reasonable; <3% expensive. Used in ValueComposite alongside Magic Formula and Acquirer's Multiple.
Composite Score
0.55 × ValueComposite + 0.45 × QualityComposite × SafetyGate. ValueComposite blends Magic Rank + Acquirer M + FCF Yield. QualityComposite blends F-Score + ROIC. SafetyGate is binary (1 if Altman Z'' ≥ 1.10 AND Beneish M ≤ -1.78, else 0 = pick rejected).
IV (Intrinsic Value)
Per-pick analyst-derived fair value (DCF or comparable-company). Long-term resolver closes the position when price reaches IV ("IV-attainment"), the thesis breaks, or the horizon expires. Drawdowns alone never trigger an exit on long-term picks.
Horizon
Long-Term Value: 1y-3y+ holding period. Swing: 1w-1m. The horizon determines which exit triggers apply (long-term: thesis/IV/horizon; swing: TP/SL/time-stop).
Sources: Piotroski 2000 ("Value Investing: The Use of Historical Financial Statement Information") · Greenblatt 2006 ("The Little Book That Beats the Market") · Carlisle 2014 ("Deep Value") · Altman 1968/2002 · Beneish 1999 ("The Detection of Earnings Manipulation"). Composite spec: findings/SYNTHESIS.md §3.
US Equity Picks
Building track record for Long-Term Value Holds. No picks emitted yet - the screener has not yet produced n≥100 closed trades.
This tab will populate once the weekly/daily screener writes its first batch of picks. See updates/long_term_value_project_2026-04-27/PROJECT.md for status.
Building track record for Swing Plays. No picks emitted yet - the screener has not yet produced n≥100 closed trades.
This tab will populate once the weekly/daily screener writes its first batch of picks. See updates/long_term_value_project_2026-04-27/PROJECT.md for status.
Building track record for Closed Holds. No picks emitted yet - the screener has not yet produced n≥100 closed trades.
This tab will populate once the weekly/daily screener writes its first batch of picks. See updates/long_term_value_project_2026-04-27/PROJECT.md for status.
Building track record · n=0/100
First weekly value-screener run scheduled. Picks will populate here once the runner emits.
See .github/workflows/value_screener_weekly.yml ·
Charter: docs/PERFORMANCE_CHARTER.md