Skip to content

research: add Hyperliquid vault strategy experiments NB01–NB23 - #61

Merged
miohtama merged 1 commit into
masterfrom
research/hyperliquid-vov-second
Oct 14, 2026
Merged

miohtama merged 1 commit into
masterfrom
research/hyperliquid-vov-second

Conversation

@miohtama

Copy link
Copy Markdown
Contributor

Hyperliquid vault strategy research, NB01–NB23

This research develops C00 from the deployed Hyper-AI rules and records the unsuccessful alternatives as well as the leads. It covers ranking, feature forecasting, position sizing, risk estimation, allocation constraints, residual cash and execution accounting.

The strongest retained benchmark is C00: four-input performance ranking, regularised 90-day risk and a capacity waterfall, with six positions. Under the latest common accounting, its 2026 replay returns 58.59% CAGR / 2.474 Sharpe / −6.36% drawdown, versus 49.15% / 2.214 / −6.41% for v11. Full-history C00 is 33.13% / 1.651 / −9.94%, versus 18.55% / 1.374 / −7.92% for v11. These are historical results on inspected, overlapping periods.

Reading the tables

  • S: 1 January–4 October 2026 inclusive, [2026-01-01, 2026-10-05) UTC.
  • L0: 1 March 2023–4 October 2026 inclusive, the original full-history period.
  • L: 1 March 2023–7 October 2026 inclusive, [2023-03-01, 2026-10-08); final archive day partial.
  • In a row marked S / L or S / L0, each metric lists the corresponding short / long result in that order. The experiment overview shows one explicitly named representative per notebook, usually its 2026 result; it is not a table of independent winners.
  • — means no applicable portfolio CAGR/Sharpe or no headline reported here. Forecast IC, overlapping fixed-unit snapshot returns and hypothetical fixed-ledger sensitivities are not executable strategy metrics.
  • Portfolio accounts start with USD 150,000. All preceding indicator history is retained, including 2025 warmup for fresh 2026 accounts. Long accounts carry their own holdings and accumulated profits into 2026.

Results come from the executed notebooks and their exact saved generations, rather than older prose in dated appendices. Accounting evolves across the research: NB01–13 retain the legacy convention; NB14–17 introduce as-of-observation NAV, with the inherited partial-sale treatment; NB18 corrects partial-sale fees; NB19–23 also reset stale entry tax and share final-proceeds funding. Early NB19 historical controls are displayed separately. Compare methods with the matched control inside their notebook. The v11 change from 41.20% to 49.15% in later controls is an accounting difference, not evidence that a new scoring method improved v11. The fee-sensitive NB17 exit-delay lead reverses after correction.

Experiment overview

One row per main notebook. The description identifies the configuration supplying the displayed metrics; complete method coverage follows below.

ID Experiment link Description Backtesting period CAGR Sharpe Maximum drawdown
NB01 Notebook Source v11 replication: 60/40 CAGR360–Sortino45, six slots, 90-day inverse variance. Exact control for the first research phase. S; L0 also run 41.20% 1.960 −6.40%
NB02 Notebook Accepted-dollar cap prevents redistribution above initial requests. Representative capped arm leaves much more cash. S; L0 also run 9.31% 1.310 −4.22%
NB03 Notebook Equal, inverse-volatility/equal blend, regularised inverse volatility and correlation risk budgets, under both allocators. Headline: correlation-aware parent. S; L0 also run 35.52% 1.620 −7.84%
NB04 Notebook Sparse metric/window matrices; clipping, missingness, blend, breadth and sizing searches. Headline: six-slot 50/50 clipping blend. S; L0 also run 41.38% 2.119 −4.85%
NB05 Notebook Hard/smooth clipping, feasible entries, replacement hurdles, tie order and cash refill. Headline: feasible-entry selector. S; L0 confirmations 41.83% 1.99 −6.39%
NB06 Notebook 74 fixed two/three/four-input score recipes. Headline: 48/32/10/10 CAGR360–Sortino45–CAGR180–Sortino30. S; L0 also run 51.77% 2.303 −6.35%
NB07 Notebook Repeat 74 scores with 6/7/8/10 positions and N/(N+2) entry/exit ranks. Headline: seven-slot Blend46d Sharpe lead. S; L0 also run 47.93% 2.459 −4.71%
NB08 Notebook 139 score forecasts; 7/30/90-day labels, learned absolute floors, percentile floors and random vetoes. Headline: unchanged focal six-slot account; learned floors veto no entries. S; L0 portfolio checks 51.77% 2.303 −6.35%
NB09 Notebook Individual feature forecasting plus six K-ratio windows, including max history; missing=True and sparse-data repairs. No portfolio backtest. S; 7/30/90-day targets — — —
NB10 Notebook Original matched allocation snapshots and expanded 36-arm smoke: risk/K inside ranking or sizing scores, five maps, breadth, covariance and missingness. S; original 7/30/90d, smoke 30d — — —
NB11 Notebook 40 fixed-basket sizing policies: 45/90/180-day risk, powers, shrinkage, multi-window and sparse observation models. No trading-engine account. S; 7/30/90-day snapshots — — —
NB12 Notebook 12 local sensitivities, eight constrained risk-budget policies and 16 engine replays. Headline: focal gap-log 90-day variance. S 47.46% 2.123 −7.10%
NB13 Notebook Ten frozen 2026 leaders plus two controls over full history. Headline: gap-log allocation CAGR leader; focal four-input score leads Sharpe. L 27.44% 1.505 −10.24%
NB14 Notebook Factorial source/regularised variance × sequential/simultaneous allocation plus downside sizing. Headline: four-input source-risk A. S; L also run 51.77% 2.304 −6.35%
NB15 Notebook A/C/G versus source v11 on matched accounts; investigate September cash and selected-basket capacity. Headline: A. S; L also run 51.77% 2.304 −6.35%
NB16 Notebook Risk-priority waterfall for v11/A/C/G. Headline: G-W fresh-2026 leader; C-W is the stronger full-history challenger. S; L also run 56.19% 2.456 −6.81%
NB17 Notebook 14 capacity, retention, overflow and optimisation policies, with partial-fee sensitivities. Headline: M05 full-history primary profit leader, before fee correction. L; S also run 35.30% 1.648 −12.12%
NB18 Notebook Protected core with risk-limited extra positions, pooled/score gates, gradual exposure and economic exits. Headline: corrected-fee C00 control. S; L also run 58.58% 2.492 −6.09%
NB19 Notebook Exact C00 core plus absolute flat/risk/score/combined tickets, hold versus resize and slot/cap sensitivities. Headline: RS01. S; L also run 58.84% 2.475 −6.45%
NB20 Notebook Clean C00/v11 comparison with partial-sale correction, entry-tax reset and final-proceeds funding. Headline: C00. S; L also run 58.59% 2.474 −6.36%
NB21 Notebook Forecast-driven marginal net-profit allocation, including conditional losses/magnitudes, loss budgets and ablations. Headline: calibrated-score adaptive arm C01. S; L shortlist 11.28% 1.654 −1.90%
NB22 Notebook Managed entry attribution and local forecast calibration. Seven-day horizon registered; pre-2026 economic gate failed, so overlay accounts were skipped. Pre-2026 qualification; S diagnostics; no overlay backtest — — —
NB23 Notebook C00 in NB160 reporting layout plus separate source v11 comparison. Four fresh accounts reproduce NB20 equity, trades and metrics. S; L also run 58.59% 2.474 −6.36%

Methods and combinations tried

This table covers the executed method families and named engine treatments. NB10 lists all 36 smoke combinations; NB11 groups its 40 neighbouring parameter policies. Grid coefficients and sensitivity cells remain in their notebooks. The 334-policy NB10 expansion and generalisation across all 139 recipes were registered but not executed. They are not counted as results. In NB10 notation, A/G/Min are arithmetic/geometric/minimum aggregations; F/L/B are focal four-input/v11/Blend46d; I/V are inverse-variance/inverse-volatility scores; C/Sh/So/Ca/D/K are CAGR/Sharpe/Sortino/Calmar/drawdown/K components. Missingness codes and constraint projections are defined in the notebook.

ID Experiment link Description Backtesting period CAGR Sharpe Maximum drawdown
01/v11 NB01 Source inverse variance S / L0 41.20% / 15.77% 1.960 / 1.238 -6.40% / -7.92%
02/cap NB02 Source inverse variance + accepted-dollar cap S / L0 9.31% / 2.08% 1.310 / 0.355 -4.22% / -7.92%
03/P1 NB03 50/50 inverse volatility plus equal, source redistribution S / L0 34.28% / 14.23% 1.485 / 1.067 -9.11% / -8.87%
03/P1-cap NB03 50/50 inverse volatility plus equal, accepted-dollar cap S / L0 24.93% / 10.39% 1.489 / 1.012 -8.17% / -8.01%
03/P2 NB03 Equal weights, source redistribution S / L0 35.01% / 13.57% 1.443 / 0.990 -9.53% / -11.22%
03/P2-cap NB03 Equal weights, accepted-dollar cap S / L0 30.10% / 11.49% 1.509 / 0.977 -9.35% / -9.29%
03/P3 NB03 Regularised inverse volatility, source redistribution S / L0 35.43% / 14.98% 1.606 / 1.140 -7.69% / -7.92%
03/P3-cap NB03 Regularised inverse volatility, accepted-dollar cap S / L0 17.40% / 8.23% 1.278 / 0.943 -6.68% / -7.92%
03/P4 NB03 Correlation-aware equal risk budgets, source redistribution S / L0 35.52% / 14.24% 1.620 / 1.081 -7.84% / -7.92%
03/P4-cap NB03 Correlation-aware equal risk budgets, accepted-dollar cap S / L0 19.30% / 8.06% 1.387 / 0.911 -6.71% / -7.92%
04/matrix NB04 Sparse CAGR/Sharpe/Sortino/drawdown/Calmar × 30/46/60/90/120/180/270/360/730d; separate CAGR/Sharpe fits, 50/50 and 60/40 score blends, 4/6/10 slots; equal, score-proportional and source variance sizing. No universal winning family. S; L0 — — —
04/blend46 NB04 Named two-input 60/40 CAGR360–Sortino46 recipe with 50/50 clipping-profile blend and source variance sizing. S / L0 41.38% / 15.42% 2.119 / 1.24 −4.85% / −8.51%
05/calibration-grid NB05 18 calibrations: hard CAGR clips 0.5/1/2 and Sortino 3/7/12, blended 45/46-day profiles, smooth saturation, ratio floors, missing-zero and an activity filter. S; selected L0 confirmations — — —
05/ties NB05 Five seeded tie-order controls, seeds 11/23/37/53/71, isolate sensitivity caused by clipping ties without selecting a new score. S — — —
05/feasible NB05 Reject deposit-infeasible entrants before incumbent replacement. S / L0 41.83% / 15.87% 1.99 / 1.25 −6.39% / −7.92%
05/hurdle NB05 Replacement improvement thresholds 0, 0.025 and 0.05, with fee/tie diagnostics; 0.05 representative. S 41.78% 1.99 −6.40%
05/smooth NB05 Smooth C1/S3 score transformation, compared with hard clips and clipping blends. S 31.64% 1.65 −5.28%
05/refill NB05 Residual cash refill after admission checks; source-score representative. S 39.56% 1.88 −6.44%
05/activity NB05 Require ten non-flat returns; explicitly tests an activity filter rather than treating flat marks as invalid. S 27.39% 1.31 −6.75%
06/three NB06 36 three-input 48/32/20 recipes with extra metric/window inputs and matched-weight controls. Highest inspected 2026 CAGR representative. S; L0 51.99% 2.29 −7.29%
06/four NB06 36 four-input 48/32/10/10 recipes; named focal uses CAGR180 and Sortino30 additions. S / L0 51.77% / 22.05% 2.303 / 1.622 −6.35% / −8.08%
06/four-peak NB06 Highest-CAGR four-input support adds Sortino180 and CAGR60 to the 48/32 core; it has worse drawdown than the focal candidate. S; L0 52.30% 2.23 −7.62%
06/collapsed NB06 Collapse added windows into a two-input 58/42 score, preserving total metric weights. S / L0 41.89% / 15.85% 1.99 / 1.24 —
07/N7 NB07 Focal four-input recipe at seven slots; entry rank 7, exit rank 9. S / L0 30.85% / 17.30% 1.582 / 1.26 −6.42% / −9.99%
07/N8 NB07 Same focal recipe at eight slots; entry rank 8, exit rank 10. S / L0 16.24% / 14.88% 1.04 / 1.12 −6.57% / −11.91%
07/N10 NB07 Same focal recipe at ten slots; entry rank 10, exit rank 12. S / L0 −3.01% / 8.49% −0.28 / 0.70 −6.79% / −17.42%
07/blend7 NB07 Blend46d at seven slots: separate higher-Sharpe candidate. S / L0 47.93% / 18.40% 2.459 / 1.395 −4.71% / −8.08%
08/forecasts NB08 139 scores; monthly absolute calibration and rank IC at 7/30/90d, quarter-coloured scatters, trend/threshold, market residual and cohort diagnostics. S; 7/30/90d targets — — —
08/absolute NB08 Monthly positive-return floors, with rejected slots held as cash or redistributed within caps. Focal six-slot rule changes no entries. S / L0 51.77% / 22.05% 2.303 / 1.622 −6.35% / −8.08%
08/percentile NB08 75th-percentile entry floors plus matched random vetoes; v11 six-slot representative has no 2026 effect. S / L0 41.20% / 16.24% 1.960 / 1.266 −6.40% / −7.92%
08/missing NB08 Reject unscored fresh entries; v11 six-slot representative. Sparse/missing scores remain a separate category. S / L0 41.20% / 11.11% 1.960 / 0.971 −6.40% / −6.39%
08/cash-yield NB08 3%/5% cash-yield overlays on fixed ledgers; descriptive account sensitivities, not executable cash strategies. S; diagnostic only — — —
09/features NB09 Individual CAGR/Sharpe/Sortino/drawdown/Calmar and activity/age/missingness controls; monthly matured-label forecasts. No account CAGR. S; 7/30/90d targets — — —
09/K NB09 K30/60/90/180/360/max ablations, missing=True, partial histories, imputation and robust scaling. Max K has a small lead; year-only addition hurts. S; 7/30/90d targets — — —
10/original NB10 Original fixed-unit snapshots: equal, inverse volatility and inverse variance; raw/peer-shrunk risk, risk windows, 4/6/8/10 slots, caps, covariance budgets and future-risk diagnostic. S; 7/30/90d snapshots — — —
10/T00 NB10 Rank L; size source90 → source; N=6, missing=Legacy, constraints=C. Exact source control. S; 30d snapshots — — —
10/T01 NB10 Rank F; size constant 1 → equal; N=6, missing=M0, constraints=C. Same-basket equal. S; 30d snapshots — — —
10/T02 NB10 Rank F; size F → linear; N=6, missing=M0, constraints=C. Frozen score sizing. S; 30d snapshots — — —
10/T03 NB10 Rank F; size I90 → linear; N=6, missing=M0, constraints=C. Inverse variance component. S; 30d snapshots — — —
10/T04 NB10 Rank F; size V90 → linear; N=6, missing=M0, constraints=C. Inverse volatility component. S; 30d snapshots — — —
10/T05 NB10 Rank F; size A(F,I90) → linear; N=6, missing=M0, constraints=C. Risk in sizing score. S; 30d snapshots — — —
10/T06 NB10 Rank A(F,I90); size constant 1 → equal; N=6, missing=M0, constraints=C. Risk in ranking only. S; 30d snapshots — — —
10/T07 NB10 Rank A(F,I90); size A(F,I90) → linear; N=6, missing=M0, constraints=C. Risk in both roles. S; 30d snapshots — — —
10/T08 NB10 Rank F; size A(F,V90) → linear; N=6, missing=M0, constraints=C. Inverse volatility blend. S; 30d snapshots — — —
10/T09 NB10 Rank F; size G(F,I90) → linear; N=6, missing=M0, constraints=C. Geometric agreement. S; 30d snapshots — — —
10/T10 NB10 Rank F; size A(F,I90) → power; N=6, missing=M0, constraints=C. Squared conviction. S; 30d snapshots — — —
10/T11 NB10 Rank F; size A(F,I90) → softmax; N=6, missing=M0, constraints=C. Softmax conviction. S; 30d snapshots — — —
10/T12 NB10 Rank F; size A(F,I90) → percentile; N=6, missing=M0, constraints=C. Ordering-based sizing. S; 30d snapshots — — —
10/T13 NB10 Rank F; size A(F,I90,K90) → linear; N=6, missing=M0, constraints=C. K in sizing. S; 30d snapshots — — —
10/T14 NB10 Rank A(F,I90,K90); size A(F,I90,K90) → linear; N=6, missing=M0, constraints=C. Three-input joint score. S; 30d snapshots — — —
10/T15 NB10 Rank A(F,I90,Kmax); size A(F,I90,Kmax) → linear; N=6, missing=M0, constraints=C. Maximum-history K. S; 30d snapshots — — —
10/T16 NB10 Rank A(F,I90,K360); size A(F,I90,K360) → linear; N=6, missing=M0, constraints=C. Long fixed K. S; 30d snapshots — — —
10/T17 NB10 Rank B; size G(B,I90,K30+) → linear; N=6, missing=M0, constraints=C. Positive K geometric blend. S; 30d snapshots — — —
10/T18 NB10 Rank L; size Min(Sh90,Ca180,V180,Kmax) → linear; N=6, missing=M0, constraints=C. Minimum and zero/fallback probe. S; 30d snapshots — — —
10/T19 NB10 Rank A(C180,So30,K90); size A(I90,Kmax,D90) → linear; N=6, missing=M0, constraints=C. Different ranking and sizing features. S; 30d snapshots — — —
10/T20 NB10 Rank A(F,I90,K90); size A(F,I90,K90) → linear; N=10, missing=M0, constraints=C. Breadth control. S; 30d snapshots — — —
10/T21 NB10 Rank A(F,I90,K90); size A(F,I90,K90) → linear; N=6, missing=MN, constraints=C. Neutral missingness. S; 30d snapshots — — —
10/T22 NB10 Rank A(F,I90,K90); size A(F,I90,K90) → linear; N=6, missing=MA, constraints=C. Available-component mean. S; 30d snapshots — — —
10/T23 NB10 Rank A(F,I90,K90_conf); size A(F,I90,K90_conf) → linear; N=6, missing=M0, constraints=C. K support confidence. S; 30d snapshots — — —
10/T24 NB10 Rank F; size A(F,I90,K90) → risk_budget; N=6, missing=M0, constraints=C. Score risk contributions. S; 30d snapshots — — —
10/T25 NB10 Rank A(F,I90,K90); size A(F,I90,K90) → linear; N=6, missing=M0, constraints=P. Uncapped mechanism control. S; 30d snapshots — — —
10/T26 NB10 Rank A(F,I90,K90); size A(F,I90) → linear; N=6, missing=M0, constraints=C. K-off sizing on K-on basket. S; 30d snapshots — — —
10/T27 NB10 Rank F; size A(F,constant 0.5) → linear; N=6, missing=M0, constraints=C. Neutral-offset flattening control. S; 30d snapshots — — —
10/T28 NB10 Rank C30; size C30 → linear; N=6, missing=M0, constraints=C. Short CAGR. S; 30d snapshots — — —
10/T29 NB10 Rank C360; size C360 → linear; N=6, missing=M0, constraints=C. Long CAGR. S; 30d snapshots — — —
10/T30 NB10 Rank Sh90; size Sh90 → linear; N=6, missing=M0, constraints=C. Short Sharpe. S; 30d snapshots — — —
10/T31 NB10 Rank Sh360; size Sh360 → linear; N=6, missing=M0, constraints=C. Long Sharpe. S; 30d snapshots — — —
10/T32 NB10 Rank So180; size So180 → linear; N=6, missing=M0, constraints=C. Sortino. S; 30d snapshots — — —
10/T33 NB10 Rank D90; size D90 → linear; N=6, missing=M0, constraints=C. Drawdown. S; 30d snapshots — — —
10/T34 NB10 Rank Ca360; size Ca360 → linear; N=6, missing=M0, constraints=C. Calmar. S; 30d snapshots — — —
10/T35 NB10 Rank Kmax; size Kmax → linear; N=6, missing=M0, constraints=C. K alone. S; 30d snapshots — — —
11/risk-grid NB11 40 policies: 45/90/180d risk; powers 1/1.5/2; κ30 or L/3; raw floor, bounded inverse-risk scores, min-component recipe, median/max multi-window risk, calendar-log and gap-log variance. S; 7/30/90d snapshots — — —
12/constraints NB12 Eight matched projected/constrained convex equal-risk or score-risk allocations, plus twelve sensitivities in 75/105d windows, κ10/90 and raw floors. Snapshot comparisons have no strategy CAGR. S; 30d snapshots — — —
12/S01 NB12 F:equal; full trading-engine shortlist replay. S 37.97% 1.526 -11.58%
12/S02 NB12 L:equal; full trading-engine shortlist replay. S 33.63% 1.376 -10.56%
12/S03 NB12 F:regular L90/p2/κ30; full trading-engine shortlist replay. S 46.60% 2.092 -7.05%
12/S04 NB12 L:regular L90/p2/κ30; full trading-engine shortlist replay. S 33.34% 1.632 -6.60%
12/S15 NB12 L:bounded_V; full trading-engine shortlist replay. S 34.25% 1.464 -9.95%
12/S16 NB12 L:T18; full trading-engine shortlist replay. S 39.66% 1.642 -9.75%
12/S17 NB12 F:calendar_log L90/p2/κ30; full trading-engine shortlist replay. S 47.04% 2.116 -7.08%
12/S18 NB12 F:gap_log L90/p2/κ30; full trading-engine shortlist replay. S 47.46% 2.123 -7.10%
12/C01 NB12 F:equal_risk / projected; full trading-engine shortlist replay. S 42.98% 1.916 -8.26%
12/C02 NB12 F:equal_risk / constrained_convex; full trading-engine shortlist replay. S 42.88% 1.925 -8.06%
12/C05 NB12 L:equal_risk / projected; full trading-engine shortlist replay. S 33.49% 1.574 -7.68%
12/C06 NB12 L:equal_risk / constrained_convex; full trading-engine shortlist replay. S 33.88% 1.602 -7.50%
12/X01 NB12 F:regular L75/p2/κ30; full trading-engine shortlist replay. S 46.92% 2.123 -7.08%
12/X08 NB12 L:regular L90/p2/κ90; full trading-engine shortlist replay. S 32.07% 1.528 -7.06%
13/shortlist NB13 Ten 2026-selected configurations replayed over full history: six score/breadth candidates and four regularised/calendar/gap-risk candidates, plus two controls. CAGR leader C07 shown; Sharpe leader C03 is 22.09% / 1.622. L 27.44% 1.505 −10.24%
14/A NB14 Four-input ranking; original source 90-day variance and sequential capacity allocation. S / L 51.77% / 22.53% 2.304 / 1.653 -6.35% / -7.92%
14/B NB14 Same source-risk basket; simultaneous capped proportional redistribution. S / L 50.46% / 22.69% 2.253 / 1.635 -6.54% / -7.92%
14/C NB14 Peer-shrunk 90-day empirical variance with partial-history support; original sequential allocator. S / L 47.56% / 27.53% 2.123 / 1.560 -6.93% / -10.73%
14/D NB14 Same regularised variance with simultaneous capped redistribution. S / L 46.60% / 27.66% 2.093 / 1.543 -7.05% / -10.63%
14/G NB14 Regularised negative-return second moment instead of total variance; original allocator. S / L 47.28% / 27.85% 2.114 / 1.543 -7.72% / -11.37%
16/V11-W NB16 Source v11 risk ordered by waterfall, preserving two-input ranking. S / L 31.57% / 15.36% 1.553 / 1.203 -6.64% / -8.66%
16/A-W NB16 Four-input score with source variance priority and waterfall. S / L 48.89% / 22.58% 2.138 / 1.647 -6.07% / -7.92%
16/C-W NB16 Four-input score with regularised variance priority and waterfall; predecessor of C00. S / L 51.35% / 30.96% 2.230 / 1.725 -6.08% / -10.33%
16/G-W NB16 Four-input score with regularised downside priority and waterfall. S / L 56.19% / 29.94% 2.456 / 1.598 -6.81% / -12.83%
17/M01 NB17 Admission-aware risk waterfall S / L 46.97% / 30.74% 2.041 / 1.713 -6.70% / -10.63%
17/M02 NB17 Proportional capped risk S / L 46.00% / 27.56% 2.064 / 1.537 -7.38% / -10.63%
17/M03 NB17 Score-priority waterfall S / L 40.99% / 28.23% 1.316 / 1.409 -15.59% / -13.81%
17/M04 NB17 Core + two score-first overflow S / L 40.33% / 28.53% 1.769 / 1.522 -7.45% / -14.04%
17/M05 NB17 Core + four score-first overflow S / L 40.13% / 35.30% 1.760 / 1.648 -7.60% / -12.12%
17/M06 NB17 Core + four scored-only overflow S / L 40.13% / 27.86% 1.760 / 1.519 -7.60% / -10.63%
17/M07 NB17 Capacity-aware incumbent retention S / L 49.18% / 29.32% 2.200 / 1.619 -6.70% / -11.01%
17/M08 NB17 Greedy score-dollar basket S / L 33.46% / 14.95% 1.131 / 0.981 -15.40% / -13.56%
17/M09 NB17 Maintained / fee-aware overflow S / L 37.61% / 27.92% 1.719 / 1.359 -8.48% / -11.79%
17/M10 NB17 One-decision rank-exit confirmation S / L 52.01% / 27.15% 2.148 / 1.515 -10.25% / -10.72%
17/M11 NB17 Joint six-slot score optimiser S / L 32.32% / 14.51% 1.101 / 0.956 -15.40% / -13.57%
17/M12 NB17 Legacy capacity extension S / L 40.30% / 32.48% 1.799 / 1.561 -8.64% / -11.24%
17/M13 NB17 Fixed ten-name core S / L -8.48% / 15.95% -0.896 / 0.824 -7.98% / -23.03%
17/fee-M00 NB17 Same M00 policy with corrected partial-redemption fees; separates accounting from allocation. S / L 58.58% / 33.13% 2.492 / 1.822 -6.09% / -9.97%
17/fee-M05 NB17 Same M05 policy with corrected partial-redemption fees; separates accounting from allocation. S / L 47.81% / 37.73% 2.044 / 1.716 -7.36% / -12.09%
17/fee-M10 NB17 Same M10 policy with corrected partial-redemption fees; separates accounting from allocation. S / L 57.17% / 28.64% 2.320 / 1.590 -10.22% / -10.41%
18/C00 NB18 Exact six-slot regularised-risk waterfall with corrected partial-sale fees. S / L 58.58% / 33.13% 2.492 / 1.822 -6.09% / -9.97%
18/C01 NB18 Check usable deposit capacity before constructing the protected six-position core. S / L 54.54% / 32.93% 2.316 / 1.811 -6.07% / -10.32%
18/C02 NB18 Spend residual cash on up to four unconditional score-ranked overflow positions. S / L 47.81% / 37.73% 2.044 / 1.716 -7.36% / -12.09%
18/R00 NB18 Additional inverse-volatility sleeve under cash, gross exposure and linear risk-headroom limits. S / L 52.46% / 31.48% 2.232 / 1.742 -6.82% / -10.32%
18/R01 NB18 Risk-only sleeve plus fee-aware hold-versus-cash exits using matured forecasts. S / L 54.09% / 31.48% 2.302 / 1.736 -6.07% / -10.32%
18/R02 NB18 Require a positive pooled lower forecast bound before additional positions are funded. S / L 54.54% / 32.50% 2.316 / 1.797 -6.07% / -10.32%
18/R03 NB18 Use positive net forecasts within score/support groups to admit additional positions. S / L 54.54% / 32.57% 2.316 / 1.792 -6.07% / -10.32%
18/R04 NB18 Scale admitted sleeve exposure gradually using expected gain and incremental risk. S / L 54.54% / 32.50% 2.316 / 1.797 -6.07% / -10.32%
18/R05 NB18 Same gradual policy with the full reference-risk budget instead of the 75% budget. S / L 54.54% / 32.23% 2.316 / 1.783 -6.07% / -10.32%
19/C00 NB19 Exact corrected C-W S / L 58.59% / 33.13% 2.474 / 1.651 -6.36% / -9.94%
19/E00 NB19 Protected C00 core; fixed additional tickets of 2.5% conservative equity, then hold units. S / L 54.53% / 33.52% 2.310 / 1.666 -7.30% / -9.94%
19/R00 NB19 Same tickets multiplied by min(1, sqrt(peer variance / own predicted variance)); hold units. S / L 54.63% / 33.41% 2.319 / 1.662 -7.29% / -9.94%
19/S00 NB19 Same tickets multiplied by 0.5 + 0.5 × clipped performance score; hold units. S / L 54.39% / 33.49% 2.308 / 1.664 -7.29% / -9.94%
19/RS00 NB19 Multiply absolute tickets by both performance-strength and risk factors; hold units. S / L 54.50% / 33.31% 2.317 / 1.658 -7.29% / -9.94%
19/RS01 NB19 Continuously resize the same absolute score × risk tickets rather than hold purchase units. S / L 58.84% / 33.66% 2.475 / 1.673 -6.45% / -9.94%
19/NRS00 NB19 Normalise score × risk across extras to spend the residual sleeve budget, with caps. S / L 56.49% / 33.05% 2.350 / 1.643 -7.29% / -9.94%
19/RS08 NB19 Combined held-ticket policy with eight total positions; C00 core ranks remain fixed. S / L 55.35% / 33.41% 2.348 / 1.663 -7.27% / -9.94%
19/RS12 NB19 Combined held-ticket policy with twelve total positions; C00 core ranks remain fixed. S / L 54.64% / 33.24% 2.320 / 1.654 -7.29% / -9.94%
19/GF25 NB19 Reduce the additional-position score floor from 0.5 to 0.25, retaining held tickets. S / L 54.42% / 33.26% 2.315 / 1.656 -7.29% / -9.94%
19/CAP10 NB19 Reduce additional gross limits from 5%/25% to 2% per name / 10% overall. S / L 57.17% / 33.00% 2.422 / 1.645 -6.72% / -9.94%
20/V11 NB20 Local v11 source with shared corrected partial sales, fresh-entry tax reset and final-proceeds funding. S / L 49.15% / 18.55% 2.214 / 1.374 -6.41% / -7.92%
20/C00 NB20 48/32/10/10 four-input ranking; regularised 90d inverse-risk waterfall; exact six-slot source selector and constraints. S / L 58.59% / 33.13% 2.474 / 1.651 -6.36% / -9.94%
21/F01 NB21 Exact C00 with sell rebalance threshold widened from 0.5% to 1%; no new forecasting. S / L 58.70% / 33.22% 2.479 / 1.664 -6.46% / -9.92%
21/P00 NB21 Adaptive marginal net-profit targets using support-conditioned pooled gain/loss distributions. S / L 10.44% / 5.68% 1.530 / 0.654 -2.01% / -14.32%
21/D00 NB21 Loss-probability feature model plus support-conditioned gain/loss magnitudes and a portfolio downside budget. S -2.73% -1.526 -2.30%
21/M00 NB21 Predict gain and loss magnitudes as well as loss probability; optimise fee-aware marginal opportunity value. S -39.74% -1.846 -32.09%
21/C01 NB21 Calibrate the C00 score into a net-return forecast; use it for adaptive portfolio targets. S / L 11.28% / 0.86% 1.654 / 0.145 -1.90% / -18.01%
21/D02 NB21 Restrict the conditional-downside allocator to at most six positions. S -2.73% -1.526 -2.30%
21/D03 NB21 Apply the downside forecast to fresh entries; ablate forecast-driven incumbent sales. S / L 3.19% / 8.62% 0.737 / 1.456 -2.86% / -4.62%
21/D04 NB21 Remove the downside-budget constraint from the conditional-downside allocator. S -17.85% -2.127 -14.55%
21/BHALF NB21 Half the registered downside budget; otherwise preserve the D00 package. S -1.24% -0.927 -1.39%
21/BDOUBLE NB21 Double that downside budget; otherwise preserve the D00 package. S -7.53% -1.303 -6.37%
21/H07 NB21 Seven-day forecast and budget package instead of thirty days; a joint horizon sensitivity. S -20.21% -1.515 -21.35%
21/MARGIN NB21 Subtract common calibration uncertainty from the marginal-value decision. S -0.08% -0.013 -1.68%
21/GATE14 NB21 Restore the source 14-day momentum gate within the forecast-driven package. S -2.73% -1.526 -2.30%
22/entry-calibration NB22 Initial-unit managed spell attribution; separate average-level/within-band calibration, 2d/7d horizon alignment, 999 identity/block controls and unresolved-outcome sensitivity. Gate failed; no overlay accounts. Pre-2026 gate; S diagnostic only — — —
23/replication NB23 Repackage exact C00 and source v11 in NB160 outline; new four-account replay matches NB20. No strategy change. S / L 58.59% / 33.13% 2.474 / 1.651 −6.36% / −9.94%

What the experiments taught us

Ranking, exposure and sizing answer different questions

Our score recipes have modest return-ranking information. NB08’s 30-day return IC is about 0.09 for both v11 and the four-input recipe, with calendar-block bands including zero and a reversal from positive Q1 to negative Q3. Historical extra windows improve the portfolio replay, but they do not demonstrate stable broad predictive superiority. Risk forecasts are more informative: NB09’s volatility ranking is around 0.74–0.77 on supported seven/thirty-day cohorts. A signal that ranks risk well still needs a separate economic argument before it controls total exposure.

What worked

The four-input ranking improved the matched source-risk account in both inspected periods. Its collapsed 58/42 two-input control lost most of that gain, suggesting the extra windows matter beyond aggregate CAGR/Sortino balance. The formula remains a historical selection from many inspected candidates.

Regularised risk plus waterfall allocation is the main construction lead. Near-flat or sparse histories can otherwise dominate inverse variance, and accepted allocations can differ sharply from requested weights. The regularised-risk waterfall improves its matched C parent in both periods and leads to C00 after shared accounting corrections. Waterfall alone is not generally superior: it hurts v11 and some source-risk comparisons. This is evidence for a particular combined policy, not proof that the 90-day window or shrinkage strength is optimal.

Explicit source parity and accounting audits materially changed our conclusions. Single charging of partial redemptions, clearing stale tax on reopened positions and respecting final net sale proceeds prevent spurious comparisons. NB23 reproduces all four NB20 equity paths and successful trade sequences. Holding current assumptions fixed is useful for research, while the 0.1% capital allowance remains an inherited assumption rather than an established venue fee.

What did not work

Accepted-dollar caps without redistribution leave about half the fresh-2026 portfolio in cash and slash CAGR from 41.20% to 9.31% in the early matched comparison. Flatter weights repair part of that failed allocator but do not beat the uncapped parent. Correlation risk budgets and constrained solvers also lose to simpler controls in the tested engine comparisons, despite some favourable snapshot outcomes.

More positions and more deployment often reduce profit. The focal recipe falls from 51.77% CAGR at six slots to −3.01% at ten. That intervention also changes entry/exit ranks and later allocations to shared winners; it is not an isolated diversification test. NB17 reduces September cash from roughly 49% to 2–5% with some methods, yet substantially lowers fresh-2026 profit. Waterfall inside six slots cannot absorb more than those vaults’ combined concentration/TVL capacity. In the September audit, capacity explains 96.9% of summed decision shortfalls, while rejected top-ups explain only 2.6%; these are deployment diagnostics, not foregone-profit estimates.

K-ratio, smoothing and admission floors supplied no demonstrated portfolio replacement. Max-history K has small incremental forecast leads, but supported uncertainty includes zero; adding the one-year feature alone worsens the tested point estimates. Learned absolute floors often act on no actual entries. Rejecting missing scores harms full-history return: absent long history is not a measured bad opportunity. Filling known NAV after inception improves metric coverage without making carried marks independent observations or reconstructing missing terminal outcomes.

Forecast-driven exposure and marginal-profit optimisation performed poorly. NB21’s best score-calibrated adaptive forecast account gives 11.28% CAGR in 2026; conditional magnitude allocation gives −39.74%. The profitable C00 cohort is often predicted negative, so increasing solver sophistication does not fix miscalibrated opportunity values. Restricting downside portfolios to six names or removing the loss budget does not rescue them.

NB22 corrects seven-day average bias from −0.91 to −0.19 percentage points, yet ranking remains weak, with IC −0.065 and roughly 49% sign accuracy. Closed pre-2026 vetoes appear to avoid USD 6,133, but the 1.09% value per resolved veto dollar loses to the 1.19% calendar-block median and reverses to −USD 4,742 when available open-position marks are included. The planned qualification gate therefore stops the overlay accounts. There is no unexecuted strategy-performance claim.

Smaller leads and important qualifications

NB19’s absolute score × risk resize sleeve adds only USD 255 in fresh 2026, with slightly worse drawdown. Role attribution assigns negative profit to extras and a positive gain to core feedback; it does not establish profitable extra positions in isolation. NB21’s 1% sell band adds about USD 111 in 2026 and USD 1,025 over full history. These are small execution leads, not significant confirmed improvements over C00.

Returns are concentrated. Realist Capital supplies about 35.4% of C00’s positive 2026 vault profit, and the largest account update is +6.46%. Its NAV and Goon Edging rise roughly 14% over two observed days while BTC/ETH fall over the same intervals. Curator and market checks help interpret the move; they do not prove a continuously executable NAV. Full-history large updates also include week-long reporting intervals. Fixed-ledger deletion or zeroing the best day measures influence, not the performance of an executable alternative portfolio.

All periods have been researched repeatedly and the long period overlaps 2026. Sparse observations, current inventory/guard applied retrospectively, incomplete historical permission clocks, resolved-label selection and terminal censoring limit generalisation. Unknown outcomes are not assumed zero, and confidence from forward-filled calendar rows is tracked separately from actual updates.

Likely future leads

  1. Freeze C00 and observe later data. Retain matched v11 and the small RS01/sell-band leads without selecting another winner from the same history. The current figures establish historical trade-offs, not an untouched confirmation.
  2. Align prediction with managed trades. Forecast net outcomes for actual executable fresh entries, with exits, re-entry, cost basis and unresolved positions accounted for. NB22’s four-day median versus 11.4-day capital-weighted life explains why a generic 30-day hold label can miss the decision’s economics; even the better aligned seven-day label remains approximate.
  3. Develop downside calibration before exposure control. Require a filter to beat identity and timing controls and retain its value under unfinished-outcome scenarios. Unsupported young/missing observations should retain an explicit fallback to the trusted core rather than automatically force cash.
  4. Refine risk and capacity jointly on fixed core rules. Use nearby risk-window/shrinkage controls, requested-versus-accepted attribution and scarce-capacity examples. Covariance diversification remains a possible targeted lead; the limited tested designs do not exhaust it. Avoid treating a solver optimum for a score surrogate as maximum executable profit.
  5. Allocate residual cash only with an evidenced marginal advantage. Preserve core selection and targets, account for slot/TVL limits and whole-account feedback, and compare extras with cash. More deployment is worthwhile only when it improves net profit at an acceptable risk trade-off.
  6. Improve observation evidence. Retain archived state when its independent clock is unknown, distinguish observation time from export/write time, audit isolated jumps and confirm real fee/permission behaviour. Keep causal fills, actual-update confidence and young-vault support explicit.

Publication contents and execution prerequisites

The PR includes the 23 main notebooks, presentation appendices, reusable Python helpers, plans, reviews and written conclusions. Numeric tables and Markdown results are retained. NB23’s 24 static charts remain embedded; bitmap payloads elsewhere and bundled runtime JavaScript are removed from the publication copy. Full local outputs remain in the original working folder.

No datasets, price archives, standalone images, engine states, parquet/CSV outputs, pickles, DuckDB files, logs or generated manifests are committed. The research .gitignore protects those files. The original working tree’s editor and notebook-runner changes are outside this PR.

These are archived research notebooks. Rerunning needs the recorded repaired archive, matching external strategies/trade-executor/trading-strategy checkouts and the local research input/cache generations referenced by the code. Those prerequisites are intentionally not bundled. Python is formatted in the PR copy as required by the repository’s PR rules; semantic AST parity was checked. Recorded execution fingerprints describe the measured pre-publication sources, so presentation formatting can invalidate byte-hash caches. Preserve the recorded results and regenerate identities/caches for a new run rather than relabelling old measurements.

Detailed experiment notes include individual hypotheses, verdicts and limitations. NB23 is the clean C00 reference; NB20 gives its matched accounting comparison.

@miohtama
miohtama merged commit 537456a into master Oct 14, 2026
1 check failed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant