PRE-REGISTERED RESEARCH LEDGER

Trading Hypotheses
& Honest Verdicts

Every strategy idea and risk filter tested on AI Price Patterns market memory and derivatives data — including the ones that failed out-of-sample or died after transaction costs.

8SHIPPED / LIVE
3NULL
7FAILS
21TOTAL TESTED

OOS VISUAL PERFORMANCE (2025-01 → 2026-11 · 1X UNLEVERED · NET OF FEES)

How the surviving edges actually compound vs. what breaks

HISTORICAL OOS · NO_TRADE

Cumulative 1x OOS Equity Curve (%)

4-coin equal-slot portfolio (BTC/ETH/SOL/ZEC), 12 bps round-trip taker+slippage

0%+25%+50%+75%+100%2025-012025-122026-11
End of OOS (2026-11-14) H_N5: +98.8% H_C2: +86.9% CROWD_FADE: +80.3% H_N4: -11.5%

Hold Horizon Scaling (OOS Net bps/trade)

Why 30h beats 12h/20h and why early cluster-abort fails

H_N8 · 4h p10/p90 Hold 40h+125.5 bps
H_N5 · p5/p95 Hold 30h+83.6 bps
H_N5 · p10/p90 Hold 30h+62.4 bps
Baseline · p5/p95 Hold 20h+55.8 bps
H_C2 · p10/p90 Hold 20h+38.9 bps
H_N5a · p5/p95 Hold 12h+30.6 bps
H_N4 · DynAbort (FAILS)-5.7 bps

1x Monthly Net vs. Max Drawdown (OOS)

Unlevered historical simulation across active rules

H_N84h p10/p90 + C2 Veto · 40h
+10.50%/moSharpe 2.26 · DD -11.5%
H_N99-Alt 4h Portfolio · 40h
+6.63%/moSharpe 2.04 · DD -10.7%
H_N5p5/p95 + C2 Veto · 30h
+4.41%/moSharpe 1.58 · DD -15.5%
H_C2p10/p90 + C2 Veto · 20h
+3.87%/moSharpe 1.29 · DD -26.0%
CROWD-FADEp5/p95 Base · 20h
+3.59%/moSharpe 1.52 · DD -12.1%
H_C3US Session + NO_CHOP
+0.66%/moSharpe 0.72 · DD N/A
PATH-RISK · Market MemoryWORKS · SHIPPED

Analogues predict how far price moves (range & level-touch probabilities)

Rule: Use memoryExactV1 + HAR volatility scaling + risk-state kNN (λ-shrinkage + Platt) to estimate range expansion and level-touch probabilities.

Evidence: Level-touch AUC 0.85–0.91 at 5 bars (beat gradient-boosted baseline); HAR vol scaling lifted range-expansion skill from 1–22% to 19–29%; risk-state memory outperformed in 7 of 12 cells.

Strict out-of-sample & independent-window controls.Updated 2026-11-12
get_path_risk · mcp-server/src/services/pathRiskService.ts Inspect live →
H_C2 · 2D ConstellationPROMISING · LIVE TEST

2D Memory Constellation Squeeze-Veto (C2_VETO_SQUEEZE) on crowd extremes

Rule: Veto contrarian p10/p90 (or p5/p95) crowd-extreme trades when NOW sits in the opposing runaway cluster AND 3-bar drift pushes deeper into it (cluster share ≥ 35%).

Evidence: Lifts 4-coin net return (H=20h, 12 bps cost) from +25.9 to +33.8 bps full (+7.9 bps vs ALL, p=0.019; 315 vetoed trades −17.8 bps) and in 2025–26 OOS (n=657) from +21.1 to +36.9 bps (+15.9 bps, p=0.021; 111 vetoed trades −56.9 bps, positive on BTC, ETH, ZEC).

HISTORICAL 1X ECONOMICS (OOS ≥2025 (4 coins))12 bps RT · NO_TRADE
+3.87%/mo1x Est. Net (+46.4%/yr)
+38.9 bpsNet / Trade
1.26Profit Factor
1.29OOS Sharpe
-26.0%Max DD (1x)
31.5/mo706 OOS trades
BTC/ETH/SOL/ZEC · 2021–26 · 12 bps RT · Dev <2025 / OOS ≥2025Updated 2026-11-15
constellation_battery.py · tools/ops/crowd_tracker.py Inspect live →
H_C3 · 2D ConstellationPROMISING · LIVE TEST

2D Memory Constellation Side-Gate & Chop-Veto (H5) on US-session momentum

Rule: At the 1h bar before US open (13:00/14:00 UTC), gate breakout direction by NOW/drift cluster (Cluster 0 for LONG, Cluster 2 for SHORT) and skip Cluster 1 (Chop).

Evidence: Cuts active sessions from 50% to 32% while lifting 4-coin Gross Sharpe 1.50→1.58 (OOS 1.47→1.69) and Net Sharpe 0.46→0.61 (OOS 0.24→0.52, or 0.69 with CORE_SCALAR; BTC+ETH+SOL Net Sharpe 0.94, p=0.029). C3_POLARIZED + NO_CHOP lifts OOS Net Sharpe to 0.72 (+3.14 bps/d).

HISTORICAL 1X ECONOMICS (OOS ≥2025 (488 US sessions, C3_POLARIZED + NO_CHOP))5.5 bps/side · NO_TRADE
+0.66%/mo1x Est. Net (+7.9%/yr)
+3.14 bps/dNet / Session
0.72OOS Sharpe
US Session 09:30–16:00 ET · 5.5 bps/side · Dev <2025 / OOS ≥2025Updated 2026-11-15
constellation_battery.py · tools/ops/momentum_tracker.py Inspect live →
CROWD-FADE · Derivatives / CrowdLIVE TEST

Fade crowd extremes (2-year p5/p95 global long/short account ratio, 20h hold)

Rule: Enter contrarian 20h position when Binance USD-M global long/short ratio crosses its expanding 2-year p5 (LONG) or p95 (SHORT), 5-minute publish lag respected.

Evidence: In-sample BTC/ETH/SOL/ZEC: +47 bps/trade net, Sharpe 1.65, max DD −15%. Out-of-sample 20 coins: +16 bps [−2, +36] (not replicated); 2025+ on new coins ≈ 0. Running live forward test (~601 trades needed for 80% power).

HISTORICAL 1X ECONOMICS (OOS ≥2025 (4 coins, p5/p95 H=20h))12 bps RT · NO_TRADE
+3.59%/mo1x Est. Net (+43.0%/yr)
+59.7 bpsNet / Trade
1.41Profit Factor
1.52OOS Sharpe
-12.1%Max DD (1x)
18.9/mo424 OOS trades
Started 2026-11-14 · 12 bps RT · next-minute openUpdated 2026-11-14
tools/ops/crowd_tracker.py Inspect live →
MOM-H1-H4 · Intraday MomentumLIVE TEST

US-session intraday noise-area breakout (H1/H2/H3/H4)

Rule: Trade 1h noise-area breakout during US equity hours (09:30–16:00 ET) on BTC+ETH; track long vs short leg (H3) and memoryExactV1 core filter (H4).

Evidence: 24 coins 2022–26: Gross Sharpe 1.25 [0.49, 1.95] (23/24 coins positive), breakeven 4.4 bps/side (6.8 BTC, 8.1 ETH). BTC+ETH next-min exec: gross 1.66, net at 5.5 bps/side 0.41 [−0.42, 1.19]. H4 core filter Fisher p ≈ 0.03 across 24+20 coins.

Started 2026-11-16 · 5.5 bps/side · ~350 sessions for powerUpdated 2026-11-16
tools/ops/momentum_tracker.py · get_intraday_momentum Inspect live →
H_N1 · 2D ConstellationFAILS

Squeeze Momentum: flip C2_VETO_SQUEEZE bars to trade WITH the squeeze

Rule: On p10/p90 (or p5/p95) crowd-extreme bars vetoed by C2_VETO_SQUEEZE, enter WITH the runaway cluster instead of staying flat.

Evidence: Lost −50.1 bps full (n=356, 95% CI [−102.7, −4.2], p=0.035; OOS −80.0 bps at 20h, −27.7 bps at 10h) on p10/p90; lost −99.6 bps full (n=192, p=0.0003; OOS −80.4 bps 20h / −84.5 bps 10h, p=0.039) on p5/p95. Staying flat (VETO) is strictly superior.

BTC/ETH/SOL/ZEC · 2021–26 · 12 bps RT · Pre-registered Nov 15Updated 2026-11-15
constellation_next_battery.py · constellation_next_battery.out
H_N2 · 2D ConstellationFAILS

Cross-market BTC Constellation lead-lag filter on alts (ETH, SOL, ZEC)

Rule: Filter altcoin crowd extremes and ETH US-session momentum using BTC’s contemporaneous 2D Constellation cluster instead of each coin’s own cluster.

Evidence: BTC_VETO on alt crowd extremes added no edge (p10/p90 full p=0.974; p5/p95 p=0.697). Gating ETH US-session momentum by BTC cluster flipped ETH negative (−20.0 bps, AnnSharpe −0.62), whereas ETH’s own C3_SIDE_GATE hit +27.9 bps (AnnSharpe +0.86) in OOS.

ETH/SOL/ZEC · 2021–26 · 12 bps RT & 5.5 bps/side · Pre-registered Nov 15Updated 2026-11-15
constellation_next_battery.py · constellation_next_battery.out
H_N3 · 2D ConstellationNULL

Causal inverse-vol sizing & 2:1 intrabar bracket on C2_VETO_SQUEEZE

Rule: Scale kept C2_VETO_SQUEEZE trades by expanding median v96 ratio clip(0.5..1.5), or exit early via conservative 2.0σ TP / 1.0σ SL 1h intrabar bracket.

Evidence: Inverse-vol sizing added no per-trade alpha (p=0.56–0.91), though OOS AnnSharpe nudged up slightly (1.48→1.67 on p5/p95). 2:1 intrabar bracket significantly underperformed fixed 20h hold (+18.4 vs +31.5 bps on p10/p90, p=0.0060; OOS +30.3 vs +52.7 bps on p5/p95).

BTC/ETH/SOL/ZEC · 2021–26 · 12 bps RT · Pre-registered Nov 15Updated 2026-11-15
constellation_next_battery.py · constellation_next_battery.out
H_C1 · 2D ConstellationNULL

2D Memory Constellation standalone 1h directional trading

Rule: Trade 1h direction whenever NOW sits in + drifts into the dominant Constellation cluster (share ≥ 40%, H=10h).

Evidence: NOW_CLUSTER hit 50.2% dev / 50.5% OOS (−11.0 / −9.1 bps net); CONSTELLATION_ALIGNED hit 50.3% dev / 52.2% OOS (−9.5 / −7.3 bps net); MTF 8h+24h agreement −14.7 bps net OOS. Direction alone does not survive 10 bps cost.

BTC/ETH/SOL/ZEC · non-overlapping 10h · 10 bps costUpdated 2026-11-15
constellation_battery.py · constellation_battery.out
DIR-NULL · Market MemoryNULL

Unfiltered kNN analogues predict standalone price direction or first extreme

Rule: Use top-K memoryExactV1 neighbours directly to forecast directional sign or whether high/low is touched first.

Evidence: Nearest-neighbour direction 48.9–50.7% on BTC 1h/15m/5m; dispersion ratio 1.02–1.06 vs random draws. Audit confirmed search works (top 2–3% similarity) — the directional null is real.

Multiple timeframes & controlsUpdated 2026-11-12
rlx-search/src/routes/candle_search.rs · localml/shape_knn.py
DERIV-ALLBAR · Derivatives / CrowdNULL

All-bar derivatives features (OI, taker ratio, top-trader ratios) predict returns

Rule: Test standard derivatives ratios across all 1h bars without conditioning on crowd extremes.

Evidence: Full battery over all bars failed Benjamini–Hochberg multiple-testing correction.

BH-corrected across feature familyUpdated 2026-11-14
research/2026-11-market-memory/deriv/deriv_battery.py
UTC-ORB-FADE · Intraday MomentumFAILS

00:00 UTC breakout, alternative levels (ORB/prev H-L), & 1m limit-order band fade

Rule: Run paper breakout rules on 00:00 UTC day, test ORB / overnight / prev-session levels, or fade bands with 2 bps maker limits.

Evidence: 00:00 UTC day: 6 sides/day, gross 0.34, deeply negative net. ORB / other levels: all net negative in 2025+. 1m limit-order fade: 80–85% TP hit rate offset by short-vol tail losers (negative net even on optimistic fills).

24 coins · 1m & 1h bars · BH-correctedUpdated 2026-11-14
research/2026-11-market-memory/breakout/levels.py & fade_1m.py
H_F1 · Derivatives / CrowdFAILS

Extreme 8h Funding Rate Fade + Constellation C2_VETO / NO_CHOP

Rule: Fade expanding 1y p10/p90 (or p5/p95) Binance USD-M fundingRate at 00/08/16 UTC settlements (H=8h, 24h), gated by C2_VETO_SQUEEZE or NO_CHOP.

Evidence: FADE_ALL_8h p10/p90 (n=5104) lost −12.4 bps net (p=0.0073; OOS −23.1 bps, p=0.0067); OOS 24h −38.1 bps (p=0.0147). C2_VETO (−13.5 bps, p=0.92) and MOM_NO_CHOP (−14.5 bps) failed to rescue it.

BTC/ETH/SOL/ZEC · 2021–26 · 12 bps RT · Pre-registered Nov 15Updated 2026-11-15
funding_oi_constellation_battery.py · funding_oi_constellation_battery.out
H_F2 · Derivatives / CrowdFAILS

Pre-settlement 2h move fade around 00/08/16 UTC funding reset

Rule: Fade a 2h pre-settlement move (>20 bps) into funding reset when funding sits in top/bottom quartile (p25/p75), gated by NO_CHOP (H=2h & 4h).

Evidence: Dev <2025 showed only +5.7 bps gross at 2h (−6.3 bps net after 12 bps RT); in 2025–26 OOS even gross disappeared (+0.6 bps gross / −11.4 bps net; −0.7 bps gross with NO_CHOP).

BTC/ETH/SOL/ZEC · n=6411 · 12 bps RT · Pre-registered Nov 15Updated 2026-11-15
funding_oi_constellation_battery.py · funding_oi_constellation_battery.out
H_F3 · Derivatives / CrowdFAILS

Crowd L/S p10/p90 confirmed by realized funding rate + C2_VETO

Rule: Require realized fundingRate to confirm crowd crowding (funding > exp_median for contrarian SHORT, < exp_median for contrarian LONG) inside p10/p90 crowd extremes.

Evidence: Looked strong in Dev (+55.3 vs +37.3 bps ALL; +58.1 bps with C2), but inverted in 2025–26 OOS (+3.2 bps, p=0.589; CONF_MED+C2 = −22.0 bps vs +15.6 bps ALL).

BTC/ETH/SOL/ZEC · 2021–26 · 12 bps RT · Pre-registered Nov 15Updated 2026-11-15
funding_oi_constellation_battery.py · funding_oi_constellation_battery.out
H_N5 · 2D ConstellationPROMISING · LIVE TEST

Extended 30h Mean-Reversion Hold vs 12h/20h on C2_VETO_SQUEEZE

Rule: Hold C2_VETO_SQUEEZE contrarian crowd-extreme trades for 30h instead of 12h or 20h (same 12 bps RT cost).

Evidence: Monotonic horizon improvement across both Dev (<2025) and OOS (≥2025): on p5/p95 OOS (n=353), net expectancy rose from +30.6 bps (12h) → +55.8 bps (20h) → +83.6 bps (30h, p=0.0344, PF 1.46, Sharpe 1.58); on p10/p90 OOS (n=706), +19.1 → +38.9 → +62.4 bps (p=0.0296, PF 1.32, Sharpe 1.55). Paired 30h–20h delta was +27.8 bps (p=0.183 OOS, p=0.051 Dev).

HISTORICAL 1X ECONOMICS (OOS ≥2025 (4 coins, p5/p95 H=30h))12 bps RT · NO_TRADE
+4.41%/mo1x Est. Net (+52.9%/yr)
+83.6 bpsNet / Trade
1.46Profit Factor
1.58OOS Sharpe
-15.5%Max DD (1x)
15.8/mo353 OOS trades
BTC/ETH/SOL/ZEC · 2021–26 · 12 bps RT · Pre-registered Nov 15Updated 2026-11-15
constellation_economics_and_next.py Inspect live →
H_N4 · 2D ConstellationFAILS

Dynamic in-trade Squeeze-Abort exit during 20h C2_VETO_SQUEEZE hold

Rule: After entering a C2_VETO_SQUEEZE contrarian trade, exit early at t+k if 1h NOW + 3-bar drift flips into the opposing runaway cluster (share ≥ 35%).

Evidence: Triggered on ~95% of trades and destroyed mean-reversion edge: −37.2 bps vs fixed 20h in Dev (p=0.0004) and −44.6 bps in OOS (p=0.0096, net −5.7 bps, PF 0.94) on p10/p90; −59.2 bps in OOS (p=0.0288, net −3.4 bps) on p5/p95. Entry veto works; intra-trade cluster abort does not.

BTC/ETH/SOL/ZEC · 2021–26 · 12 bps RT · Pre-registered Nov 15Updated 2026-11-15
constellation_economics_and_next.py
H_N6 · 2D ConstellationNULL

Dual-scale 4h + 1h Constellation Gate on C2_VETO_SQUEEZE

Rule: Additionally veto C2_VETO_SQUEEZE entries when the 4h resampled Constellation NOW cluster sits in the opposing runaway cluster (share ≥ 35%).

Evidence: Added no edge: −4.7 bps in Dev (p=0.445) and −20.1 bps in OOS (p=0.074, net +18.9 bps vs +38.9 bps baseline) on p10/p90; −7.5 bps in OOS (p=0.669) on p5/p95. 1h Constellation already captures the squeeze risk.

BTC/ETH/SOL/ZEC · 2021–26 · 12 bps RT · Pre-registered Nov 15Updated 2026-11-15
constellation_economics_and_next.py
H_N7 · 2D ConstellationNULL

High Constellation Polarization (≥ dev median) + 30h Hold on C2_VETO_SQUEEZE

Rule: Require 2D Memory Constellation cluster polarization (med_0 − med_2) ≥ dev median on C2_VETO_SQUEEZE (H=30h, 12 bps RT).

Evidence: 4-coin OOS looked strong (+158.2 bps, p=0.0016 on p10/p90; +119.2 bps, p=0.046 on p5/p95), but leave-ZEC-out control on BTC+ETH+SOL failed to replicate (+9.8 bps, p=0.835 on p5/p95; +34.9 bps, p=0.339 on p10/p90). Concentration in ZECUSDT drove the 4-coin lift.

BTC/ETH/SOL/ZEC vs BTC/ETH/SOL control · OOS ≥2025 · 12 bps RTUpdated 2026-11-15
crowd_30h_refinements.py
H_N8 · 2D ConstellationPROMISING · LIVE TEST

4h Constellation Grid + 40h/48h Hold on Crowd-Fade p10/p90 + C2_VETO_SQUEEZE

Rule: Evaluate p10/p90 crowd extremes on 4h bar closes gated by native 4h 2D Memory Constellation C2_VETO_SQUEEZE (H=8, K=12) and hold for 10 bars (40h) or 12 bars (48h) at 12 bps RT.

Evidence: Monotonic improvement to 40h–48h before decaying at 60h (OOS Sharpe 1.71). At 4h 40h (n=411 OOS), net +125.5 bps (p=0.0032, PF 1.75, DD 11.50%, Sharpe +2.26; Dev +75.2 bps, p=0.0008); at 48h (n=369 OOS), +144.1 bps (p=0.0008, DD 9.49%, Sharpe +2.24). Leave-ZEC-out (BTC+ETH+SOL) at 40h holds +45.0 bps OOS (Sharpe +1.07) and stays positive in all 3 six-month OOS windows (2025-H1 +44.8, 2025-H2 +36.8, 2026 +54.2 bps; ETH +93.8, SOL +89.1, ZEC +264.2, BTC −4.0 bps).

HISTORICAL 1X ECONOMICS (OOS ≥2025 (4 coins, 4h p10/p90 H=40h))12 bps RT · NO_TRADE
+10.50%/mo1x Est. Net (+126.0%/yr)
+125.5 bpsNet / Trade
1.75Profit Factor
2.26OOS Sharpe
-11.5%Max DD (1x)
18.3/mo411 OOS trades
BTC/ETH/SOL/ZEC + BTC/ETH/SOL NO-ZEC control · 2021–26 · 12 bps RTUpdated 2026-11-15
verify_4h_horizon_curve.py Inspect live →
H_N9 · 2D ConstellationPROMISING · LIVE TEST

9-Alt H_N8 Portfolio Expansion (ETH/SOL/ZEC + XRP/DOGE/BNB/AVAX/LINK/ADA, 4h 40h)

Rule: Run H_N8 (4h close p10/p90 crowd extreme + 4h C2_VETO_SQUEEZE, 40h hold, 12 bps RT) across 9 liquid alts (excluding BTC) with ~6-slot equal-weight allocation.

Evidence: Replicates cleanly on 6 untouched alts (2023–26, 48h: Dev +59.6 bps p=0.030, OOS +62.1 bps p=0.057, DD 8.44%, Sharpe +1.40; 6/6 coins ≥0 OOS led by LINK +157.0 & DOGE +101.9 bps). Combined ALTS-9 (40h uncapped, n=730 OOS = 32.6/mo): +79.3 bps (p=0.0040, PF 1.41, DD 10.66%, Sharpe +2.04, 0 zero-trade months, max drought 15.7d). NO-ZEC-8 control holds +49.7 bps OOS (n=607, p=0.0424, DD 8.94%, Sharpe +1.46).

HISTORICAL 1X ECONOMICS (OOS ≥2025 (9 alts, 4h p10/p90 H=40h))12 bps RT · NO_TRADE
+6.63%/mo1x Est. Net (+79.6%/yr)
+79.3 bpsNet / Trade
1.41Profit Factor
2.04OOS Sharpe
-10.7%Max DD (1x)
32.6/mo730 OOS trades
9 alts + NO-ZEC-8 control · 2023–26 · 12 bps RT · Dev 23–24 / OOS ≥25Updated 2026-11-16
hn8_portfolio_risk_sim.py · tools/ops/crowd_tracker.py Inspect live →

ANTI-OVERFIT PROTOCOL

Why we publish the nulls & failures too.

Every idea is pre-registered with explicit dev (< 2025) and time-holdout OOS (≥ 2025) splits, 5-minute Binance archive publish lag, next-minute open execution, day-clustered bootstrap confidence intervals, and realistic round-trip taker + slippage costs (11–12 bps).