Skip to content

research: open hyperliquid-lower-vol track with NB01 baseline - #60

Merged
miohtama merged 133 commits into
masterfrom
research/hyperliquid-lower-vol
Sep 29, 2026
Merged

miohtama merged 133 commits into
masterfrom
research/hyperliquid-lower-vol

Conversation

@miohtama

@miohtama miohtama commented Sep 9, 2026 •

Copy link
Copy Markdown
Contributor

Summary

The scratchpad/hyperliquid-lower-vol/ research track: 43 notebooks that try to lower the volatility of the Hyperliquid vault-of-vaults incumbent (hyper-ai.py v6) without giving up its return, plus the live-versus-backtest reconciliation that closes the track's current phase.

The anchor is 02-better-format.ipynb — the incumbent on the full window 2026-01-01 to 2026-09-08 (126 two-day decisions, $150,000): CAGR 37.90%, cycle Sharpe 2.160, max drawdown −4.45%, 96 positions over 33 vaults. Every verdict below is against that anchor, under the standing gates in RESEARCH-RULES.md, with a 0.25 indifference band on cycle Sharpe (126 decisions cannot resolve less).

Nothing has been adopted. One mechanism sits just inside the indifference band; everything else is rejected, has no effect, or is a diagnostic.

The track

# Notebooks Question Verdict
0 01, 02 reproduce the release candidate; retire the frozen vault archive for the standard downloader; extend the window to 2026-09-08 anchor established
1 03a, 03b, 12, 13 how good is the data, how much power is there, what does an age barrier hide? measurement: minimum detectable Sharpe gap ≈ 2.5; marks are weekly before April 2026
2 04–07, 09, 11 portfolio volatility target, pool caps, breadth and concentration, drawdown sizing, consistency selection all REJECT
3 14–19 evidence-weighted selection and sizing, core/satellite sleeves, against a placebo frontier all REJECT
4 20–26 volatility-matched family, joint downside selection, Sortino leg swap, drop decomposition all REJECT; 20 a mark-quality diagnostic
5 28–33 do stability signals predict forward stability, and do they help as prefilters? 12 of 13 signals never reached a backtest; the screen's return clause was later found unpassable
6 34–37 exclude the volatility tail before ranking; the same as a threshold; rankers, sizers, breadth the track's best result, measured_8, is inside the band — NOT CONFIRMED
7 38–39 do trimmed trailing returns or Sharpe rank vaults better? best rank correlation 0.25 in event time, 0.14 post-April — not a usable ranker
8 40–42 which trades made the money, and can the collapses be exited faster? one position is 46% of net P&L; the collapse exits already fill at the unmoved open; a tighter gate and a one-day cadence are worse
9 43 + 44-live-parity-plan.md why does the live vault not track the backtest? same return on the live dates, four times the drawdown; a parity plan follows

Where the track ended up

41-research-trade-forensics.ipynb explains why nothing helps: the incumbent's profit is in long holds (positions held over 30 days made $37.0k of $36.7k net), its 85 rank-churn exits are worth −$0.9k in total, and 46% of net P&L is a single position. There is no volatility to remove without removing the return.

43-research-live-vs-backtest.ipynb then compared the live Lagoon vault (live since 2026-03-30, NAV $50k) with the anchor on the same dates:

anchor backtest live share price
2026-03-30 → 2026-09-08 +14.3% +14.0%
max drawdown −4.5% −16.0%
volatility / Sharpe on the two-day grid 18% / 1.76 46% / 0.85
2026-08-12 → 2026-09-08 (incumbent logic live) +2.7%, dd −2.9% −2.8%, dd −9.0%

63% of the anchor's 86.7% total return is January to March, before the vault existed. On the live dates the two end in the same place by different paths, and the paths diverge for reasons outside the alpha model: the backtest's vault universe excludes vaults that are closed to deposits today (so three vaults the live executor held are missing from its whole history), the live data is 3–61 hours stale on 5 of 18 decisions, the fixed $75/$750 trade thresholds are 0.5% of equity at the backtest's $150k but 1.4–2.5% at live's $30–55k, and the two clocks were a day out of phase for the first week.

44-live-parity-plan.md (Draft 2, reviewed by Grok) is the plan to fix that: a one-step replay that seeds the engine with the live book and equity at each live decision and scores the six vaults it intends against the six live intended, with a pre-registered acceptance test. Strategy fixes — rank hysteresis, capacity-aware selection — are staged behind it, because a backtest that cannot reproduce the live book cannot be used to evaluate a fix to it.

🤖 Generated with Claude Code

…e release candidate

Reproduces 94-candidate-fixed-perf-fee from hyperliquid-waterfall-rc as
01-initial.ipynb. The machine-specific frozen vault archive is retired in
favour of the standard Trading Strategy website downloader, and the backtest
window is extended to 2026-09-08.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
@miohtama

miohtama commented Sep 9, 2026

Copy link
Copy Markdown
Contributor Author

Live hyper-ai snapshot, 2026-09-09 10:01 UTC

Source: https://hyper-ai.tradingstrategy.ai/state, executor cycle 115. Total equity $51,851, cash $1,903, realised P&amp;L all-time $-8,311. P&L below is executed cost vs mark or proceeds; it does not deduct the redemption fees the backtest models.

Open positions

# Vault Opened Days Invested Value Unrealised %
514 Octavious 2026-08-22 17 $13,259 $12,731 $-528 -3.98%
515 Sequoia Hy 2026-08-22 17 $2,619 $2,681 $+62 +2.36%
518 Citadel 2026-08-29 10 $16,849 $16,850 $+2 +0.01%
522 Gucky_4coi 2026-09-05 3 $3,481 $3,483 $+2 +0.06%
523 Mad Scient 2026-09-09 0 $11,865 $11,825 $-41 -0.34%
524 DOEZOE 2026-09-09 0 $2,158 $2,378 $+220 +10.18%
Total $50,231 $49,948 $-283 -0.56%

Recently closed positions (last 20 with executed trades)

# Vault Opened Closed Days Invested Proceeds Realised %
516 Buyback Sp 2026-08-25 2026-09-09 14 $6,481 $6,519 $+37 +0.57%
519 AceVault H 2026-08-31 2026-09-09 8 $17,454 $16,432 $-1,022 -5.85%
520 Mad Scient 2026-09-03 2026-09-07 3 $8,639 $8,586 $-53 -0.61%
509 Goon Edgin 2026-08-20 2026-09-03 14 $17,507 $16,126 $-1,381 -7.89%
517 Gucky_4coi 2026-08-27 2026-09-01 4 $4,604 $4,609 $+6 +0.12%
506 AceVault H 2026-08-16 2026-08-29 13 $16,742 $18,451 $+1,709 +10.21%
494 Realist Ca 2026-07-28 2026-08-27 29 $5,175 $5,019 $-156 -3.01%
498 Citadel 2026-08-03 2026-08-27 23 $17,722 $18,490 $+767 +4.33%
502 Mad Scient 2026-08-08 2026-08-27 19 $9,108 $9,232 $+124 +1.36%
495 Danny Ocea 2026-07-30 2026-08-25 26 $10,742 $10,646 $-97 -0.90%
508 pmalt 2026-08-20 2026-08-25 5 $6,840 $6,453 $-387 -5.66%
507 HYPErQuant 2026-08-18 2026-08-20 2 $4,451 $4,453 $+2 +0.05%
504 Gucky_1coi 2026-08-14 2026-08-18 4 $9,887 $9,795 $-92 -0.93%
505 Satori Qua 2026-08-14 2026-08-16 2 $7,381 $7,379 $-1 -0.02%
503 22Cap 2026-08-09 2026-08-14 5 $10,384 $10,425 $+41 +0.39%
476 Scared Mon 2026-07-20 2026-08-14 24 $5,005 $4,960 $-46 -0.91%
501 Goon Edgin 2026-08-07 2026-08-09 2 $5,249 $5,079 $-170 -3.24%
500 pmalt 2026-08-05 2026-08-08 2 $6,188 $6,027 $-160 -2.59%
499 Satori Qua 2026-08-05 2026-08-07 1 $5,517 $5,495 $-22 -0.39%
470 Goon Edgin 2026-07-18 2026-08-05 18 $7,993 $7,864 $-128 -1.61%

All-time: 108 closed positions with executed trades (410 further position records never filled and are excluded), 26 winners (24%), net realised $-8,311. Median hold 7 days.

All-time realised P&L by vault (closed positions)

Vault Positions Win rate Capital cycled Realised Mean hold (d)
Goon Edgin 7 14% $36,416 $-2,134 20
-VectorZer 2 0% $699 $-699 2
Super Moon 2 0% $699 $-699 2
AJ Pro 3 33% $2,033 $-699 3
pmalt 4 0% $16,560 $-644 5
Growi HF 2 0% $604 $-604 2
Total Rebo 1 0% $604 $-593 6
22Cap 4 50% $12,346 $-540 4
Liquidator 4 50% $3,110 $-523 19
Realist Ca 7 43% $8,793 $-325 15
YEELON 4 0% $3,460 $-306 20
Titan Vaul 2 0% $5,752 $-276 39
Octavious 4 0% $1,702 $-257 6
Crypto Pla 1 0% $4,797 $-219 108
DOEZOE 2 0% $1,544 $-199 6
IKAGI 1 0% $5,061 $-119 66
WDA MCLS 1 0% $5,377 $-111
Danny Ocea 1 0% $10,742 $-97 26
Gucky_1coi 3 67% $28,624 $-75 3
Scared Mon 1 0% $5,005 $-46 24
69 Jump St 1 0% $4,624 $-41 112
L/S Moment 1 0% $5,876 $-40 10
Satori Qua 4 0% $13,602 $-32 5
Loop Fund 2 0% $4,578 $-28 60
Citadel 4 25% $27,964 $-20 8
Growi_test 4 50% $3,032 $-18 9
JizzJazz 1 0% $13 $-13 1
The Good S 1 0% $1,335 $-13 7
Equinox · 1 0% $7 $-7 13
RedCoast 1 0% $7 $-7 12
LowRiskCry 1 0% $1,335 $-5 7
kusa 2 0% $305 $-4 5
Hyperliqui 4 0% $5,541 $-3 27
Gucky_2coi 1 0% $1,335 $-3 7
LONG quali 1 0% $13 $-2 1
Jay Pennie 1 0% $13 $-2 1
OIB 1 0% $1,330 $-2 4
Ethos Inte 1 0% $1,335 $-2 7
HyperProfi 1 0% $1,743 $-1 1
FuturAI La 1 0% $7 $-0 12
HYPErQuant 1 100% $4,451 $+2 2
Hypernova 1 100% $1,335 $+3 7
Gucky_4coi 1 100% $4,604 $+6 4
Hyperdash 2 100% $1,342 $+10 10
[ Systemic 3 33% $2,673 $+11 9
Momentum E 1 100% $1,335 $+17 7
Buyback Sp 1 100% $6,481 $+37 14
Mad Scient 3 67% $23,648 $+150 10
Wild Money 1 100% $3,886 $+198 23
AceVault H 4 25% $35,520 $+660 8

@miohtama

miohtama commented Sep 9, 2026

Copy link
Copy Markdown
Contributor Author

Allocation formula assessment: separating pump vaults from steady vaults

What the live book says

The live strategy is doing exactly what the release candidate was optimised to do, and that is the problem. The backtest edge was Realist Capital, which NB88 measured at about 36 percentage points of CAGR. That vault has since blown up. Its trailing 30-day return is now -67% with a -68% drawdown. Live, the strategy has 108 filled closed positions, a 24% win rate, a 7-day median hold, and -$8,311 realised. Only AceVault, Wild Money and Mad Scientist are net positive by vault.

The metrics that separate the pump vaults from the steady ones are already visible on the current book. All figures are trailing to 2026-09-08, computed from the cached vault price history and Binance BTC via fetch_binance_price().

Vault Status 180d return 90d vol BTC beta 90d Best-5-day share of 180d return Ulcer 180d Positive 30d windows
Sequoia open +58% 202% 3.84 1.02 0.48 55%
DOEZOE open +125% 221% 2.06 1.35 0.25 43%
Octavious open +84% 70% 1.03 1.13 0.16 55%
Citadel open +13% 15% 0.06 1.07 0.010 69%
HYPErQuant closed after 2 days +36% 13% 0.03 0.46 0.011 97%
22Cap closed after 5 days +13% 4% 0.01 0.44 0.004 93%

A best-5-day share above 1.0 means every other day in the window was net negative. Octavious, DOEZOE and Sequoia are BTC leverage with a handful of good days. HYPErQuant and 22Cap are the steady compounders we want, and the current score ranked them low enough that they were dropped within days.

Two structural points explain why the score ends up here. The 360-day CAGR leg is dominated by a single +116% number, and a 45-day Sortino on an upside-only run is enormous. Then the 33% concentration cap binds on Citadel, and because allocation_pct forces 98% deployment, the overflow has to land on the next-ranked names, which are the pumpers.

On the ensemble-of-lookbacks idea

This chain already tested it, and it failed hard. NB78 replaced the composite with an average of the same score over 90, 180 and 360 days. The strict variant fell from 32% to 17% CAGR with drawdown going from -9% to -22%. Averaging windows is not a steadiness measure. A pump vault still averages high because one window carries it.

The version of the idea that is untested is a minimum across windows rather than a mean. Score a vault by its worst window Sharpe, or require it to be positive on 30, 90, 180 and 360 days simultaneously. That is a consistency requirement, which is what "steadily making profit" actually means. It is worth one notebook, screened first by the precision-at-6 test from NB47 so it does not cost a full backtest if it is bad at the top six.

Proposed changes

Change the objective before the score. Every prior pass optimised Sharpe or CAGR on a window where pumps paid, and the optimiser will find pumps again with any new score. The lower-vol track should optimise something like CAGR subject to a realised-vol ceiling, or the Martin ratio, which is return over the ulcer index. NB42 and NB43 are instructive here. Both the BTC-beta filter and the smoothness tiebreak cut drawdown from -7.8% to about -5% and were rejected only because the adoption rule was Sharpe improvement. Under a vol-targeting objective those are candidates, not failures.

Add a portfolio vol target with cash as a position. Scale total deployment so ex-ante portfolio vol sits near a target, say 15 to 20% annualised, and let the remainder sit in cash or managed yield. This is the single lever that does not depend on selection quality and that directly removes the overflow-into-pumpers mechanism.

Size by 1 over ulcer index instead of 1 over variance. Inverse variance treats upside-only volatility as risk in the same way as drawdown volatility. Ulcer only penalises time spent under water. Pair it with a quality gate, because NB77 showed inverse-variance also overweights a quiet cohort of small losers.

Selection features that measure the right property. Ranked by how well they separate the two populations above, and by what NB78 found on downside features:

  • BTC-residual return over 180 days rather than raw CAGR, using the Binance series. NB47 found residualisation hurt under a Sharpe objective because directional vaults paid. Under a lower-vol objective penalising beta is the intent.
  • Best-5-day share of trailing return, as a cross-sectional rank. It is the cleanest single-pump detector in the table.
  • Ulcer index and share of positive rolling 30-day windows.
  • A down-day share band. NB78 found the blow-ups have almost no down days at all, so too few down days is a warning, not a virtue. Gucky_4coi at 8% down days is the one open position to watch on that basis.

Two warnings from the chain that still apply. First, NB79 showed that in a 6-vault book, vetoes that removed only losers still hurt because the freed capital went further down the ranking. Any filter must be paired with the option to hold cash, or it will substitute into something worse. Second, NB51 concluded two disciplined passes over 35 features found no selection signal beating the composite for Sharpe, and recommended walk-forward validation over more feature search. The live book is now that walk-forward, and it is failing, which is the strongest argument for changing the objective rather than adding legs.

Proposed notebook order for this track

  1. 02 objective change to vol-constrained CAGR or Martin ratio
  2. 03 portfolio vol target with cash as a position
  3. 04 ulcer-index sizing
  4. 05 BTC-residual return and best-5-day-share selection, screened by precision-at-6 first
  5. 06 min-across-windows consistency score

🤖 Generated with Claude Code

miohtama and others added 2 commits September 9, 2026 12:16
Reformats 01-initial onto the waterfall-rc NB02 chart stack, ending at
Positions at the end. Strategy configuration is unchanged and every headline
figure reproduces 01-initial exactly. Adds two chart-only universe count
indicators required by the standard universe chart.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…l track

Plan for NB03a-NB11: reserve a hold-out up front, measure NAV staleness and
the 2026-04 regime break, fix capacity before breadth, then test vol
targeting, drawdown-based sizing, a residual-event concentration penalty and
a min-across-windows consistency score under a constrained objective.
Includes the Codex CLI review the second draft responds to.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
@miohtama

miohtama commented Sep 9, 2026

Copy link
Copy Markdown
Contributor Author

Smoothing experiment plan added

The largest changes the review forced: a hold-out (2026-07-01 to 2026-09-08) reserved before any variant runs, a data-quality notebook first because NB57 measured 47 to 57% stale vault days in February and March with a polling-regime break in April, capacity realism ahead of breadth, and a constrained objective (CAGR, volatility, ulcer, invested-basket beta, time in market) ranked by Martin ratio in place of a standalone smoothness score.

miohtama and others added 10 commits September 9, 2026 12:38
Hook map for the three baseline cells, a shared run_variant and panel
harness built on the existing MASKED_VAULTS mechanism for leave-one-vault-out,
and per-notebook indicator, sizing and decide_trades code for NB03a to NB11,
written against the actual baseline so other agents can build the notebooks.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
… track

Confirms the 2026-04 polling regime break on fresh data (57.6% stale
vault-days Jan-Mar vs 38.1% Apr-Jun), and finds the development window's
return is earned by low-beta vaults (67.8% of net P&L from |beta|<0.2,
-9.9% from |beta|>0.6) rather than by the high-beta pumpers that motivated
the track - those enter the live book in July-September, inside the
reserved hold-out. NB08-NB10's case has to be made on the hold-out and the
live-book study, not a development-window improvement. Detectable Sharpe
difference on this window is about 1.60.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
min_window_sortino and positive_window_share clear the precision-at-6 gate
in both polling regimes and proceed to NB09; residual_cagr_score fails
(confirming NB47 under a Martin-ratio objective), deferring NB10;
residual_event_concentration fails as a standalone ranker, downgrading
NB08 to a diagnostic run inside NB09 rather than its own sweep.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Every target_portfolio_vol from 10% to 20% either misses the 30% CAGR
floor or ends up with a higher invested-basket BTC beta than the anchor,
because the anchor is already low-beta in this window (NB03a). Every
target also lowers the Martin ratio, since CAGR falls faster than the
ulcer index as deployment shrinks uniformly. No configuration carried to
NB11's development-window combination; worth a hold-out diagnostic where
the BTC-beta pumpers this lever targets actually occur.

Also commits the notebook builder scaffolding (_build/) used to generate
NB03a onward: shared indicator/parameter additions, the enhanced
decide_trades/compute_sizing_weights, and the run_variant/panel harness.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
…axes)

Every per_position_cap_of_pool_pct tighter than the baseline 33% lowers
CAGR sharply (29.2%/20.2%/7.9% at 15%/10%/5%) and raises the ulcer index
at 10% and 5%, because tightening the cap forces substitution into
shallower vaults (mean discarded-for-liquidity value rises from $41.9k to
$193.0k at 5%) rather than removing capital from the book - the same
mechanism NB79 documented for vetoes. Baseline 33% cap kept unchanged and
carried forward to NB06.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
… a near-miss)

Breadth at 8 or 10 vaults collapses CAGR (18.3%/4.8% vs anchor's 48.8%)
and Martin ratio. Concentration at 25% is close - CAGR and ulcer both
improve slightly - but fails on invested-basket beta (0.0181 vs anchor's
0.0142); 20% is worse on every axis. Anchor's 6-vault/33%-concentration
configuration kept unchanged. Flags 25% concentration as worth combining
with a selection-side change (NB09) in NB11.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
inverse_ulcer and inverse_downside both worsen CAGR and the ulcer index
relative to inverse_variance (38.97%/2.29% and 38.08%/2.18% vs the
anchor's 48.76%/1.40%), the opposite of the NB77 hypothesis this notebook
tested. The beta-group cap is a no-op at 40%/50% (anchor already holds
almost no high-beta vaults) and a narrow miss at 25% (fails only on
volatility, by 0.15pp). Anchor sizing kept unchanged; this closes out
every non-selection lever tried in NB04-NB07, all REJECT.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
…8 lead

min_window_sortino and positive_window_share both fail badly as full
selection legs despite clearing NB03b's precision-at-6 screen (CAGR
22.4%/34.2% vs the anchor's 48.8%, ulcer roughly doubling), collapsing
hardest in the sparse polling regime - the screen and the full backtest
are different tests. The deferred NB08 diagnostic (event-concentration
penalty at lambda=0.5) is the strongest result in the track: CAGR 55.3%,
ulcer 1.32% (both better than anchor), Martin ratio 41.9, failing only
narrowly on volatility and beta. Flagged as a single-point lead for NB11,
not adopted (no plateau, no leave-one-vault-out yet). NB10 not built
(residual_cagr_score failed NB03b's gate).

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
… problem

The NB09 event-concentration lead does not survive a plateau check
(Martin ratio 32.8/42.0/28.0 at lambda 0.25/0.5/0.75 - a spike, not a
plateau, per NB79's rule); nothing from NB04-NB09 is carried to the
hold-out. The anchor's own fresh hold-out run (2026-07-01 to 2026-09-08)
confirms the track's premise starkly: CAGR falls from 48.8% to 18.6% and
Martin ratio from 34.8 to 9.4. A fresh hold-out deployment converges to
the identical closing basket as the full-window backtest and the live
hyper-ai executor (AceVault, Citadel, DOEZOE, Mad Scientists, Octavious,
Sequoia) regardless of starting point. Track closes with no adopted
change; recommends untested vault-feed metadata, a circuit-breaker
mechanism, or an operational response as next steps.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
@miohtama

miohtama commented Sep 9, 2026

Copy link
Copy Markdown
Contributor Author

Smoothing experiment plan executed: NB03a-NB11

Ran the full 03-smoothing-experiment-plan.md track. Every notebook is committed with real backtest output, not placeholders.

Notebook Verdict
NB03a data quality, attribution, power Anchor's dev-window return comes from low-beta vaults (67.8% of net P&L from |beta|<0.2); the pumpers arrive only in July-Sept
NB03b feature screen min_window_sortino and positive_window_share pass; residual_cagr_score fails (NB10 not built)
NB04 vol target REJECT - anchor already low-beta here, so scaling deployment down only hurts CAGR
NB05 pool cap REJECT - tightening forces substitution into thinner vaults, hurting both return and smoothness
NB06 breadth/concentration REJECT - breadth collapses CAGR; 25% concentration a near-miss on beta alone
NB07 drawdown sizing REJECT - inverse_ulcer/inverse_downside both worse than inverse_variance here
NB08 event-concentration penalty Not built as its own sweep (failed NB03b's gate as a standalone ranker); tested as a single diagnostic inside NB09
NB09 consistency selection REJECT - both features that passed the screen collapse in the full backtest, worst in the stale-data regime
NB10 residual CAGR leg Not built - failed NB03b's gate, confirming the original chain's NB47 finding
NB11 close-out The one lead (NB09's diagnostic) fails a plateau check - a spike, not a real effect

The real finding is in NB11's hold-out. The anchor's own performance, run fresh with no January-June carryover, falls from 48.76% CAGR / Martin ratio 34.77 on the development window to 18.59% CAGR / Martin ratio 9.45 on 2026-07-01 to 2026-09-08 - exactly where the BTC-beta pumpers live. A fresh deployment there converges to the same six vaults as both the full-window backtest and the live hyper-ai executor (AceVault, Citadel, DOEZOE, Mad Scientists, Octavious, Sequoia), regardless of starting point. That rules out path-dependency as the explanation: the composite mechanistically finds these vaults from the data itself.

No change is adopted. Every lever this track tried - volatility targeting, pool-cap tightening, breadth, concentration, drawdown-based sizing, two consistency scores, and one penalty - was rejected on the development window or failed the plateau check. NB11 recommends three directions not yet tried: vault-feed metadata independent of share-price staleness (leader stake, commission, account P&L), an explicit circuit-breaker reacting to live BTC moves rather than trailing statistics, or an operational response (smaller sizes, faster exits) rather than a further allocation-formula change.

🤖 Generated with Claude Code

miohtama and others added 2 commits September 9, 2026 16:56
Runs the original and enhanced decide_trades side by side in one kernel,
with a same-code control run to separate logic differences from run-order
effects.

Verified: the enhanced decide_trades is behaviourally identical to the
original with all feature flags off - identical 417 trades matched on
pair, direction, timestamp and value, identical final equity, drawdown
and position count.

Found two measurement defects the track's verdicts partly rest on:
- refresh_vault_redemption_accounting mutates sell_tax on the shared
  strategy_universe, so the first backtest in a kernel sits on a
  different equity path than later ones (ann vol +0.85% relative). Every
  notebook runs its anchor first and variants after, so the bias is
  systematic.
- panel() resamples a 2-day-cycle equity curve to daily and zero-fills,
  making time_in_market measure cadence (0.497) rather than deployment
  (actual invested fraction 0.975), and attenuating invested-basket beta
  by about 3.1x and its R-squared by about 7.8x.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Fixes found by _build/verify-anchor-parity.ipynb:

1. Backtests in one kernel were not independent.
   refresh_vault_redemption_accounting mutates sell_tax on the shared
   strategy_universe, so the first run in a kernel sat on a different
   mark-to-market path than later ones. Since every notebook runs its
   anchor first and variants after, the bias was systematic. run_variant
   now resets the fee before each run; two runs of identical code are now
   bit-identical (control gap $428 -> $0.00).

2. The panel measured a 2-day-cycle strategy on a daily clock. Resampling
   to daily and zero-filling made time_in_market report cadence (0.49)
   rather than deployment (0.975), and attenuated invested-basket beta by
   3.2x and its R-squared by 8.4x. Volatility, Sharpe, Sortino and beta now
   use the strategy's own cycle clock, time_in_market is replaced by
   mean_invested, and beta uses a full-window regression against BTC
   compounded over matching periods. Also caps weights at
   max_concentration_pct before the NB04 ex-ante vol estimate, guards
   luck_ratio against the undefined opposite-sign case, and fixes NB03a's
   power calculation to the cycle clock.

All verdicts are unchanged - every lever still REJECT, the NB09 lead is
still a spike, the hold-out still degrades sharply - but several findings
are now better founded, and two were artefacts:
- NB07's beta-group cap is an exact no-op at all three thresholds
  (identical to the anchor, paired CI [0.000, 0.000]); its earlier
  'narrow miss at 25%' was contamination.
- NB09's strict vs permissive scoring is an exact no-op, not merely close.
- NB04 now shows the cash-overlay effect it was meant to catch: deployment
  falls to 61-85% against the anchor's 97.5%.
- NB11's hold-out beta is now computable (0.0830, above the development
  window's 0.0711) where the 90-day rolling estimate returned NaN.

Also guards builder.write_notebook against resetting written-up headings
back to placeholders on rebuild.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
@miohtama

miohtama commented Sep 9, 2026

Copy link
Copy Markdown
Contributor Author

Review found two harness defects; both fixed, NB03a-NB11 re-run

Reviewed the track by running the original and modified decide_trades side by side in one kernel (_build/verify-anchor-parity.ipynb) rather than by inspection.

The core check passes. The modified decide_trades reproduces the original 01-initial strategy exactly with all feature flags off: 417 trades matched trade-by-trade on pair, direction, timestamp and executed value, and identical final equity, drawdown and position count.

Two defects found and fixed:

  1. Backtests in one kernel were not independent. refresh_vault_redemption_accounting mutates sell_tax on the shared universe, so the first run sat on a different mark-to-market path than later ones. Every notebook runs its anchor first and variants after, so the bias was systematic. Proven by control: two runs of identical code differed by $428. Now reset before each run, control gap $0.00.
  2. The panel measured a 2-day-cycle strategy on a daily clock. Zero-filling off-cycle days made time_in_market report cadence (0.49) instead of deployment (0.975), and attenuated invested-basket BTC beta 3.2x and its R-squared 8.4x. Now computed on the strategy's own cycle clock, with mean_invested from portfolio statistics and a full-window beta regression against BTC compounded over matching periods.

Also fixed: NB04's ex-ante volatility now caps weights at the concentration ceiling; luck_ratio returns NaN instead of a misleading negative when the legs have opposite signs; NB03a's power calculation moved to the cycle clock.

All verdicts are unchanged — every lever still REJECT, the NB09 lead is still a spike, the hold-out still degrades sharply — but two earlier findings were artefacts and several are now better founded:

Change Before After
NB07 beta-group cap "narrow miss at 25%" Exact no-op at all three thresholds, paired CI [0.000, 0.000]. Never binds, so untested rather than rejected
NB09 strict vs permissive "nearly identical" Exactly identical; the distinction does not exist on this universe
NB04 cash overlay invisible (0.49 for every target) Deployment falls to 61-85% vs anchor 97.5% — the effect the metric was meant to catch
NB11 hold-out beta NaN (90-day window on a 70-day period) 0.0830, above the development window's 0.0711
Anchor beta 0.0142 0.0711
NB03a detectable Sharpe difference 1.60 1.68

One further nuance the corrected beta makes visible: breadth (NB06) and both consistency legs (NB09) genuinely do cut BTC exposure — beta falls to 0.017-0.032 against the anchor's 0.071 — but destroy so much return that they fail the CAGR floor. They were rejected for failing Goal 1, not Goal 2.

🤖 Generated with Claude Code

miohtama and others added 2 commits September 9, 2026 17:44
Recomputes NB03b's precision-at-6 screen at both alignments. NB03b reads
features with .asof(when), returning the bar labelled when, which per
NB57 is left-labelled and closes at when+1d - one bar ahead of what
decide_trades sees via get_indicator_value (index=-1).

The look-ahead changes one gate verdict: downside_deviation_90 fails as
published (sparse -0.240) but passes at live parity (sparse +1.139), so
it should have earned a selection backtest and did not. The decisive
verdicts are unaffected: residual_cagr_score still fails (NB10 stands),
residual_event_concentration still fails (NB08 downgrade stands), and
min_window_sortino and positive_window_share still pass.

Also quantifies a second screen defect: the mean top-6 basket holds only
4.64 of 6 names, and unequally across features (ulcer_index_180 1.2,
positive_window_share 3.3, incumbent 6.0), so the screen is not a
like-for-like comparison - which weakens positive_window_share's apparent
winning margin in particular.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
All eight items from the second review, verified in place:

1. NB03b's feature screen read .asof(when) - the bar labelled when, which
   is left-labelled and closes a day later - giving it information
   decide_trades cannot see. Now reads one bar back.
2. downside_deviation_90, which the alignment fix promotes from fail to
   pass, gets the selection backtest it earned (cagr_downside_weight).
3. The screen scored unequal baskets (4.64 of 6, and 1.2 vs 6.0 across
   features). It now scores full baskets only and reports the fill rate.
4. The pre-registered vol-matched placebo, omitted in the first pass, now
   runs.
5. The NB57 Stage C vault-feed features the plan listed are now screened.
6. NB11's basket comparison matched truncated symbols against full names
   and would have silently reported the opposite of the truth; it now
   matches on pool address.
7. NB11's verdict table is derived from the plateau result computed in the
   same notebook instead of hardcoded.
8. NB05 gains the TVL-floor sweep the plan asked for and a corrected
   column name; NB07 gains equal-risk-contribution sizing with a
   residual-correlation cap, the family the plan specified; NB03a's
   availability audit now checks raw poll data; the paired bootstrap moved
   to the cycle clock.

Findings that changed:

- Three Stage C features (cumulative_volume, follower_count,
  leader_commission) clear the gate. They are independent of the share
  price and so free of NAV staleness - the best-supported next step.
- positive_window_share's apparent win was an unequal-basket artefact; on
  full baskets it cannot score the sparse regime at all.
- The vol-matched placebo - a control meant to be a null - satisfies every
  adoption constraint at three contiguous settings (N=25/30/35, Martin
  39.25 vs anchor 34.77), the only configuration in the track to do so. It
  then fails leave-one-vault-out and is not carried to the hold-out. It
  also shows the consistency legs' beta reduction is generic de-risking,
  reproducing NB42.
- NB05's TVL floor at 25,000 passes all five constraints but has a lower
  Martin ratio than the anchor and fails leave-one-vault-out.
- NB03a's audit confirms observation-time alignment but shows written_at
  is ~10 days later due to bulk data repair, so the archive cannot prove
  point-in-time availability - a data limitation affecting the whole chain.

Verdicts are otherwise unchanged: no lever adopted, nothing reaches the
hold-out, and the anchor still degrades from 48.76% to 19.20% CAGR there.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
@miohtama

miohtama commented Sep 9, 2026

Copy link
Copy Markdown
Contributor Author

All eight review fixes applied, NB03a-NB11 re-run and verified

Every item from the fix list is in place and checked programmatically against the executed notebooks.

# Fix Verified
1 Screen reads one bar back, not the bar that closes a day later code uses when - ONE_BAR, no .asof(when) remains
2 downside_deviation_90 gets the selection backtest the alignment fix earned it cagr_downside_weight arm runs in NB09
3 Screen scores full baskets only, fill rate reported full_basket_rate column present
4 Pre-registered vol-matched placebo now runs seven settings swept in NB09
5 NB57 Stage C vault-feed features screened all five in the NB03b table
6 NB11 basket compared on pool address, not truncated symbol correctly reports 3 of 3 pumpers
7 NB11 verdict table derived from the plateau computed in-notebook no longer hardcoded
8 NB05 TVL-floor sweep + column rename; NB07 equal-risk-contribution sizing; NB03a audit against raw polls; bootstrap on the cycle clock all present in outputs

What the fixes changed

The control became the candidate, then the robustness rule killed it. The vol-matched placebo — drop the N highest-volatility candidates, no view on quality — satisfies every adoption constraint at three contiguous settings (N=25/30/35), a genuine plateau, Martin ratio 39.25 against the anchor's 34.77. It is the only configuration in this track ever to clear the bar. Leave-one-vault-out then drops it to 32.13% CAGR with the ulcer index rising above the anchor's, so it is one vault deep and does not reach the hold-out. It also proves the consistency legs' beta reduction was generic de-risking, reproducing NB42's finding.

Three Stage C features clear the gate: cumulative_volume, follower_count, leader_commission. They are measured independently of the share price, so they escape the NAV staleness that undermines every price-derived statistic here. Screening them was this track's job; testing them as selection legs is now an evidence-backed next step rather than a guess.

positive_window_share's apparent win was an artefact. On equal baskets it cannot score the sparse regime at all — its earlier margin came from a 3-name basket being compared against the incumbent's 6.

NB05's TVL floor at 25,000 passes all five constraints (CAGR 32.9%, volatility 10.6% vs 15.2%, beta 0.045 vs 0.071) but has a lower Martin ratio than the anchor and also fails leave-one-vault-out.

NB03a's audit found a data limitation affecting the whole chain: observation-time alignment is correct, but written_at sits ~10 days after the poll because rows were rewritten by a bulk repair pass, so this archive cannot prove point-in-time availability — only observation-time alignment.

Verdicts are otherwise unchanged. No lever is adopted, nothing reaches the hold-out, and the anchor still degrades from 48.76% to 19.20% CAGR there while converging on all three known pumpers.

🤖 Generated with Claude Code

miohtama and others added 3 commits September 9, 2026 18:47
Re-runs all 22 configurations from NB04-NB09 plus the anchor in one
kernel and puts them on a common axis: equity curves (combined and per
family), underwater curves, a CAGR-versus-ulcer trade-off scatter with
BTC beta as marker size, a combined metrics table, and a failure-mode
breakdown. Every figure reproduces its source notebook exactly.

Findings only a combined view gives:
- The ulcer index is the most commonly failed constraint (12 of 22).
  Most variants made the curve rougher, not smoother - they gave up
  return and got more drawdown, rather than trading one for the other.
- Only two variants beat the anchor on both return and smoothness:
  the event-concentration penalty (a spike NB11 rejected) and
  vol_matched_drop_25 (the crudest rule in the track).
- Families fail by characteristic mode, not degree: sizing and
  concentration variants all fail volatility+ulcer+beta together;
  consistency legs fail CAGR+volatility+ulcer; vol targets miss on a
  single constraint each.
- All four configurations that pass the five constraints are exclusion
  rules that shrink the candidate pool before ranking, not changes to
  how candidates are scored or sized - which cuts against the plan's
  premise that the allocation formula was the thing to fix.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Every notebook (NB03a-NB12) now ends at 2026-09-08, matching NB01 and
NB02. The development/hold-out split is retired: the period NB11
previously reserved is now inside the data every variant is selected on,
so the track carries no out-of-sample claim. The formerly-reserved period
is still reported as late_cagr / late_ulcer sub-period columns and NB11
still runs it as a fresh deployment, but it can only describe the
deterioration now, not validate against it.

Two verdicts changed, both against the candidate:

- The vol-matched rule passed at N=25/30/35 on the shorter window, a
  plateau. On the full window only 30 and 35 pass, with 30 sitting 15
  Martin points above both neighbours - a spike - and it then fails
  leave-one-vault-out. A plateau is only as stable as the window it is
  computed on.
- min_tvl_usd=25,000 cleared all five constraints on the shorter window
  and misses the CAGR floor on the full one.

The binding constraint changed too: the CAGR floor is now the most
commonly failed (13 of 22 variants) where the ulcer index was before,
because the later period is where the strategy earns least. NB03b's gate
now passes five features, four of them NB57 Stage C metadata measured
independently of the share price.

Verdict is unchanged: nothing is adopted.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
`cagr_lookback_days = 360` acts as an undocumented minimum-age rule: 225 of 336
vaults in the trading universe (67%) can never be scored, an unscored vault ranks
at signal 0 and never wins a basket slot, and the youngest vault the anchor ever
bought was 361 days old at entry. All 43 post-2026-04-01 launches sit behind it.

The cohort behind the barrier is worse, not better: median life Sharpe -0.43
against +0.14 for the scorable cohort. At tradable size only 2 of 32 have a
Sharpe resolvable at t > 2, and the one qualifying post-April name earns 27.2%
CAGR, below the track's 30% floor. A decision-aligned screen of the minimal fix
(90-day CAGR leg) loses the dense regime by 4.00 pp of mean 30-day forward
return and fails the both-regimes gate.

Verdict: nothing adopted, no NB14 warranted.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
@miohtama

miohtama commented Sep 9, 2026

Copy link
Copy Markdown
Contributor Author

NB13 - are we missing high-Sharpe vaults because of our inclusion rules?

Question asked: are there high-Sharpe Hyperliquid vaults the strategy never considers because of an inclusion rule rather than because it ranked them badly, particularly vaults launched after 2026-04-01?

Short answer: one, and it fails our own return floor.

Notebook: 13-research-age-barrier.ipynb (commit 1aee045). Descriptive only, nothing adopted.

There is an age rule, and nobody wrote it

The strategy has no explicit age filter. But cagr_lookback_days = 360 means the CAGR leg of the selection composite is NaN until a vault has 360 days of price history, an unscored vault enters the ranking at signal 0.0, and so it sorts below every scored vault and never wins a basket slot.

Measured on the run's own indicator series:

Input Parameter Median lag after inception Never valid
cagr_score cagr_lookback_days = 360 360 d 210 (62.5%)
sortino_score sharpe_lookback_days = 45 45 d 13 (3.9%)
cagr_sortino_weight both legs 360 d 225 (67.0%)
inverse_vol inverse_vol_window = 90 90 d 15 (4.5%)
return_gate gate_lookback_days = 14 14 d 2 (0.6%)

225 of the 336 vaults in the trading universe (67%) can never be scored at any point in the window. All 43 post-2026-04-01 launches are among them.

It binds completely. Across all 96 anchor positions the youngest vault ever bought was 361 days old at entry, mean 512 days. Not one entry below the line.

The curator is not responsible: build_hyperliquid_vault_universe() runs with min_age=0.0 and admits 339 of 601 Hypercore vaults with no age test. Its exclusions are deposit-closed (191), peak TVL below $7,500 (48), denylist / sub-vault / risk (23).

Is anything good behind the barrier?

Mostly not. Median life Sharpe is -0.43 for the hidden cohort against +0.14 for the scorable one; median life CAGR -24.5% against -19.0%. Of the 32 hidden vaults at tradable size ($50k TVL, 45+ fresh marks), only 2 have a Sharpe distinguishable from zero at t > 2, and only one of those also beats the anchor's 37.90% CAGR - Cold Process, which launched in February and is not part of the post-April cohort.

The five post-2026-04-01 vaults large enough to hold, over their own life to 2026-09-08:

Vault Inception Fresh days Sharpe t CAGR Max DD Last TVL
Stratwise Multi-Asset Public 2026-07-16 49 7.67 2.70 27.2% -0.5% $308k
Quantitative Market Neutral Strategy 2026-04-15 144 1.92 1.20 24.8% -4.5% $974k
Abs Return Sophon Quant 2026-05-11 120 -0.36 -0.21 -3.5% -5.7% $145k
korea wins again 2026-06-25 75 -2.47 -1.11 -87.3% -47.5% $131k
NEET WORLD ORDER 2026-05-26 87 -2.90 -1.41 -97.8% -68.6% $277k

Stratwise is the only resolvable positive, and its 27.2% CAGR sits below the track's own 30% floor and 10.7 pp below the anchor. The Market Neutral vault has exactly the profile this track has been looking for - steady, $974k of capacity, -4.5% max drawdown - but on 144 observations its Sharpe is not distinguishable from zero.

These are not NB78 fakes, they are unresolved

NB78's loser signature is a high Sharpe manufactured by an absence of down days. That is not what is happening here. The hidden cohort's top-quartile names post down days at 0.337 of days against the scorable cohort's 0.346, and are markedly better marked (stale share 0.037 against 0.160). What separates them is statistical power: median sharpe_t of 1.22 against 2.35.

A short history does not make a Sharpe dishonest. It makes it unresolvable, which is a different and more honest reason to distrust it.

Would fixing it help? No, on current evidence

Decision-aligned screen (NB47 precision-at-6, features read at when - 1 bar for live parity per NB57, full baskets only), capacity-filtered to vaults that could absorb a full 1/6 share of a 98%-deployed book ($74,242 of TVL), 30-day forward horizon, 81 decision dates:

Rule Mean return (dense) Mean return (sparse) Median Martin (dense) Median Martin (sparse)
incumbent (360d CAGR leg), top 6 +1.42% -0.50% 0.839 -0.276
short composite (90d CAGR leg), top 6 -2.58% -0.94% -0.786 -0.220
incumbent (360d CAGR leg), top 3 +0.05% +0.02% 0.409 -0.264
hidden cohort only (90d CAGR leg), top 3 -3.13% +0.87% -0.516 +0.009

Shortening the CAGR leg to 90 days - the minimal change that would let this cohort be ranked at all - wins one cell of four and loses the dense regime, the period the question was about, by 4.00 pp of mean 30-day forward return. It fails the both-regimes gate. No NB14 follows.

Caveats

  • The sparse regime carries 15 decision dates to the dense regime's 66 (the screen skips a 60-day cold start). Every cell where the shortened window looks better is a sparse-regime cell, and none survives into the larger regime.
  • 81 dates at 2-day spacing with a 30-day forward horizon is roughly 8 non-overlapping observations per regime. NB03a put this window's minimum detectable effect well above what a screen this size resolves, so the verdict rests on the sign being consistently against the change in the larger regime, not on the magnitudes.
  • sharpe_t treats fresh observations as independent and vault marks are autocorrelated, so true errors are wider. That biases the test towards admitting hidden vaults, and even so almost none qualify.
  • A second barrier is unaddressed. inverse_vol needs 90 days, and under inverse_variance sizing a NaN becomes weight 0. Shortening only the CAGR leg would let young vaults be selected but not funded. Any real fix has to move both, which is a larger change than the one screened here.
  • The 191 deposit-closed exclusions are a separate question this notebook does not answer. A vault that closes deposits after a strong run is a different kind of missing name than a young one.

Takeaway

The barrier is undocumented and much larger than anyone intended - two thirds of the universe - but on this window it is paying for itself. Worth re-measuring if the universe's age mix shifts, since the Hyperliquid vault population is young and getting younger.

miohtama and others added 3 commits September 9, 2026 22:17
Plan to allocate more to vaults with evidenced consistent profit (Stratwise
Multi-Asset Public, Quantitative Market Neutral Strategy as exemplars of the
type) and less to single-event BTC-beta pumpers, accepting a lower CAGR
floor (20%) in exchange for a Sharpe objective, per the operator's explicit
instruction.

Draft 1 was reviewed by Codex CLI (gpt-5.6-terra) and independently
smoke-tested before any real notebook was built. The smoke test found a
serious bug in the naive Sortino t-statistic admission score: it correlated
-0.41 with realised P&L across a 75-position test run, with the single
worst loss (-$28,990 in 6 days) entered at the score's maximum on 45
observations - a thin, smooth sample scores like strong evidence right
before it can reverse. The Codex review independently flagged the same
statistic's inconsistent sampling, plus a placebo-constraint NaN bug that
defaulted an out-of-range candidate to an automatic pass, a core/satellite
sleeve that could evict hold-protected positions, and a screen (NB14) that
ranked a different score than the backtest (NB16) would trade.

Draft 2 fixes all of these: an event-time, autocorrelation-discounted,
cross-sectionally-shrunk admission score (sortino_shrunk_score) replaces
the naive statistic; the objective is reframed as Sharpe non-inferiority
plus a material ulcer improvement rather than a fragile Sharpe race; the
placebo frontier is reduced to its non-dominated envelope and fails rather
than passes outside its range; the core/satellite sleeve reserves
hold-protected positions first and records realised (not just target)
sleeve weights; and a family-wise reality check is added to NB19's
close-out. Re-running the exact smoke test that found Draft 1's bug moved
the correlation to +0.11 and the same run's P&L from -$12,983 to +$13,821.

Every splice anchor, indicator, and constraint was verified by building and
executing a real notebook against the actual codebase (_build/verify-plan14-
draft2.ipynb), not just written and reviewed on paper: anchor parity holds
exactly (CAGR 37.90%, 578 trades, $186,746), all new indicators populate,
and the inverse_vol_early early-availability fallback is proven - and
checked empirically, 0 violations - to never fire before inverse_vol itself
does at default parameters, so the anchor path is unaffected.

NB14-NB19 not yet built; this commit is the reviewed, verified plan and
shared build-system additions (_build/blocks_evidence.py,
_build/harness_evidence.py) that every subsequent notebook splices in.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
The pure evidence score (sortino_shrunk, no CAGR leg) is the only rule of
four that passes the descriptive gate, beating the incumbent composite on
mean forward return in both polling regimes and median forward Martin in
the dense regime, while reaching 35.2% of its top-6 slots into vaults
younger than the incumbent's 360-day cutoff (incumbent: 0.0%).

Both composites that blend in the evidence-shrunk CAGR leg fail the gate
badly (median dense Martin 0.078 and 0.008 against the incumbent's 0.839),
complicating the assumption inherited from NB31 that a CAGR leg is
generically load-bearing - it may be load-bearing for the incumbent's own
360-day leg specifically, not for a shrunk one.

Descriptive gate only, per the plan: every score is still backtested in
NB16 regardless of this result.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
vol_matched_drop_30 reproduces NB12's spike on a fresh, independently
downloaded snapshot: Sharpe 2.747 against the anchor's 2.160, at lower
volatility (0.1492 vs 0.1543). NB12 already showed this exact configuration
fails leave-one-vault-out, so this is further confirmation it is noise
recurring in the same place rather than a real structural improvement.

Consequence: the non-dominated envelope's own maximum volatility sits below
the anchor's, so the anchor itself fails constraint 7 (mechanically correct
under Draft 2's NaN-off-range-fails fix, not a bug) - dominated by a control
point already known to be fragile. Flagged as a caution for NB16-NB18: any
candidate landing near vol 0.149 faces an inflated placebo bar (~2.85
Sharpe) set by this spike, and should be checked against the adjacent
envelope points (drop_50 at 2.015, drop_60 at 1.472) before being read as a
clean rejection.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
miohtama and others added 27 commits September 18, 2026 12:48
…on is the 2026 result, N=4 is that vault at twice the weight, the quality floor removes the vaults that earn, thr175 passes the standing gates on the refined grid

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…bution, capped unlimited comparator, filtered uncapped six-name neighbour, two-sided plateau only, survivor-based floor population with in-trade agreement, realised sleeve check, softened claims

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…the refined grid and capped family, measured_8 and thr150 as distinct rules, aligned attribution as contribution not cause, gate prose generated from the gate strings

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…cribed as capped-weight not a sleeve, concentration wording, scoped opening summary, coverage wording

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…ed, N=4 and floor claims narrowed, floor-and-sleeve not separated, exploratory wording throughout, agreement check scoped

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…tity claims narrowed to measured shares

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…ld sentences narrowed, gate-7 runs named

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…sue remains

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…ult price paths, rebalancing split; NB40 engine vault positions in floor runs listed in full

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
…for review

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…e check (both collapses have full 4h coverage; pmalt is a grind at 4h)

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
… gate-threshold test at the incumbent's cadence, a fill kill-switch, a universe-wide forward contrast, one breaker and one cluster run at most; 4-hour arm removed

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…efined on fill price, the forward table gates nothing, the local verdict and the breaker run are removed, stop rules are real

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…dicate on planned_mid_price vs the decision bar's open, H0 first in the heading, increment as the two date windows' cycle P&L, first-entry strike table

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…iew - H0 split into price kind and crash leg, async-aware predicate, first-strike table on UTC days with the aftermath row stated, gate 6 unscripted

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…d at the unmoved open on both days (partial catches); gate -12% catches the 19 Aug day and is worse overall (sells winners' noise); the one-day clock loses $12.5k to churn and fails gate 7

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
… churn classified from the in-trade pool (tighter-gate runs by their own threshold), pre-decision fire counts, feed-delay fail-closed, causal claims narrowed

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…ses the same fail-closed checks before the stop rule, async flag agreement, in-pool sell example by the exact rule, first-strike cohort stated

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…ledger, 4-hour coverage over the whole interval, settlement override asserted as configuration

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…cket caveat on the live fill, window A one-day Sharpe on the two-day grid, H3 forced-exit fills and lock-up census, stop-table reconciliation, wording

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…return on the live dates, four times the drawdown; sixth-slot churn, extra live vaults, fill timing -$365

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…ne-step replay on intended books, stale T-1 rows, closed-vault universe

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…d the same $75/$750 constants as live, so the mismatch is the equity they apply to; NB40 manifest re-run rows

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
@miohtama

Copy link
Copy Markdown
Contributor Author

The most promising leads, and what they did to CAGR, Sharpe and drawdown

All against the anchor (02-better-format.ipynb): CAGR 37.9%, cycle Sharpe 2.160, max drawdown −4.45%, cycle volatility 15.4%, 2026-01-01 to 2026-09-08, $150,000, 126 two-day decisions. A Sharpe difference under 0.25 is inside the indifference band — this window cannot resolve it — so "NOT CONFIRMED" means conditionally positive, not adopted.

Leads that survive the standing gates

Lead Mechanism CAGR Sharpe Max dd Vol Sharpe gap Verdict
anchor — 37.9% 2.160 −4.45% 15.4% — incumbent
measured_8 (NB35, NB36) drop the 8 most volatile measurable candidates before ranking 41.0% 2.374 −4.45% 14.9% +0.214 NOT CONFIRMED
thr175 (NB37, NB40) the same as a threshold: exclude candidates with trailing volatility above 1.75 38.9% 2.294 −4.45% 14.8% +0.134 NOT CONFIRMED
thr150 the same at 1.5 — identical exclusion set, identical numbers 38.9% 2.294 −4.45% 14.8% +0.134 REJECT (gate 6: 1.25 is a cliff, not a plateau)

measured_8 is the best result the track produced: +3.1 pp of CAGR and +0.21 of Sharpe at the same drawdown, holding 32 vaults instead of 33. It passes standing gates 1, 2 (0.86 mask retention), 3, 5, 6 and 7. It is inside the band, its null diagnostic ranks it 4 of 20 random exclusions, and the diversification tie-break is unresolved — better on three measures, level on one, worse on one. The threshold form (thr175) is the same mechanism with a tunable knob and a smaller gain; thr150 and thr175 exclude the same vaults, but only thr175 sits on a plateau.

Leads that looked better and are not

Lead CAGR Sharpe Max dd Vol Why it fails
thr150_n4 63.7% 2.686 −9.7% 19.0% gate 3 (held-book volatility 0.0200 vs the anchor's 0.0184) and gate 6 (every reduction of N is a spike). NB40: it is the anchor's names at ~2× weight
nofilter_n4 54.5% 2.300 −9.7% 19.8% same two gates
n4cap 53.1% 2.265 −8.5% 19.7% same two gates
thr150_nall_invvar (hold everything, inverse variance) 9.1% 1.819 −1.13% 4.9% gate 7: the late sub-period is negative. The drawdown is four times smaller and so is the return
gate12_2d (NB42, momentum gate −12% instead of −16%) 35.5% 1.993 −5.05% 15.9% passes every standing gate at −0.167 — NOT CONFIRMED, and worse on all three headline numbers
anchor_1d (one-day cadence, NB42) 25.1% 1.461 −5.42% 16.3% gate 7: late CAGR −1.4%. On the anchor's own two-day grid its Sharpe is 1.53 against 2.16
anchor_f10 (quality floor at Sharpe 1.0 + cash sleeve, NB40) 0.6% 0.114 −12.6% 14.2% gates 6 and 7. The incumbent's winners score at the floor, and the floor has no hysteresis

The pattern across the whole track: everything that raises Sharpe materially does so by concentrating — fewer names, higher weights — and then fails the held-book volatility gate or the plateau. Everything that genuinely lowers volatility (thr100 at 20.6% CAGR / 9.8% vol, the hold-everything sizer at 4.9% vol) gives up most of the return with it.

The caveat that applies to all of them

NB40 measured the overlap: every six-name lead is the anchor with one or two positions moved — 98–99% capital overlap — and one position (20 June to 21 August) is 46% of the anchor's net P&L. Under a single-vault mask the anchor retains 0.825 of its Sharpe and thr175 0.853. These are not independent strategies; they are the same book, and the leads inherit its dependence on one name.

NB43 adds the live check: on the live dates the anchor returns +14.3% and the live vault +14.0%, so the backtest is not optimistic about the level — but its drawdown is −4.5% against live's −16.0%, and the live book differs from the backtest's at the sixth slot on half the decisions. Until 44-live-parity-plan.md closes that gap, promoting any of these leads would be tuning a model of a portfolio we are not running.

🤖 Generated with Claude Code

@miohtama
miohtama merged commit 0be3392 into master Sep 29, 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