Repository navigation
research: add Hyperliquid vault strategy experiments NB01–NB23 - #61
Merged
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
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
[2026-01-01, 2026-10-05)UTC.[2023-03-01, 2026-10-08); final archive day partial.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.
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.
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
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
.gitignoreprotects 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-strategycheckouts 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.