research: add Lighter vault-of-vaults backtests NB01-NB02

[miohtama]

What

Ports the Hyperliquid vault-of-vaults research chain to the Lighter (ZK-rollup perps DEX) public-pool universe — synthetic chain 9998, 34 USDC pools, synthetic lighter-pool-{account_index} addresses. Restricting to USDC also excludes Lighter's separate USDG-denominated Robinhood Chain deployment published under the same synthetic chain id.

01-lighter-initial.ipynb

Replicates hyperliquid-waterfall-rc/17-backtest-overspent-fix.ipynb on Lighter: positive trailing-180-day-Sharpe gate, 0.875-year age ramp, 20-pool basket, 20% concentration and pool-TVL caps, synchronous-cash cap, 100,000 USD initial capital, 2025-08-01 → 2026-07-10.

Metric	Result
Cumulative return / CAGR	20.60% / 22.12%
Sharpe / Sortino	0.70 / 1.12
Max drawdown	-26.65%
Trades / positions	1,985 / 99
Mean accepted deployment / mean cash	80.4% / 21.5%

The synchronous-cash cap completes without OutOfSimulatedBalance. Capacity is the binding constraint: the young, thin cohort leaves ~1.9 signals per cycle pool-size-capped.

02-backtest-rebalancing-method-trials.ipynb

Reruns all 31 catalogued rebalancing methods (NB22–NB158) on the same Lighter universe, mirroring hyperliquid-waterfall-rc/04-backtest-rebalancing-method-trials.ipynb. All 31 complete. Best holdout Sharpe is 1.49 (NB101, age ramp + 30d trend overlay), with the NB158 waterfall baseline a co-leader at 1.48 and the highest holdout CAGR (73.72%). Method ranking inverts relative to Hyperliquid — NB28 (diversified waterfall), the Hyperliquid headline winner, is the single worst method on Lighter.

The dominant robustness caveat: neutralising each method's top 5% of profitable days turns all 31 methods negative over the full period, so the entire Lighter result is tail-driven. Full analysis in the notebook headings; a venue-comparison analysis of why Lighter underperforms Hyperliquid follows as a PR comment.

Dependencies

Needs the Lighter universe support merged today: trade-executor #1572 with companions trading-strategy #241 and web3-ethereum-defi #1348 (all merged).

🤖 Generated with Claude Code

[miohtama]

Why does Lighter underperform Hyperliquid — weaker vaults, or something else?

A first-pass, strategy-free comparison of the two raw vault populations (current metrics feed, same eligibility floors: peak TVL ≥ $7.5k, deposits open) says the raw Lighter pools are not weaker — their medians are slightly better:

Population stat	Hypercore (n=351)	Lighter USDC (n=34)
Median age	0.67y	0.73y
Median lifetime CAGR	−19.8%	−8.5%
Share with positive lifetime CAGR	35%	44%
Median 3M Sharpe	−0.01	0.56
Median current/peak TVL	48%	32%

Both populations are majority-losing; both venues' profits live in a winning tail. So the backtest gap (3.8 holdout Sharpe on Hyperliquid vs ~1.5 on Lighter) needs another explanation.

Suspected drivers, ranked

1. Universe curation asymmetry (prime suspect). The Hyperliquid backtests do not run on the raw population — build_hyperliquid_vault_universe() takes the top 120 of ~350+ eligible vaults sorted by trailing 1-year CAGR. That is recent-winner pre-selection baked into the universe before the strategy runs (disclosed in the HL trial plan as static-membership survivorship bias). The Lighter universe is all 34 eligible pools — there is nothing to cut, so its numbers are more honest. Part of the gap is likely that the Hyperliquid numbers are inflated, not that Lighter is deficient.
2. Breadth / selectivity. A 20-vault basket from a curated 120 picks the top ~17% of a winner-tilted pool. A 20-pool basket from 34 is 59% of the universe — selection has almost no room to avoid the losing majority. Evidence: in NB01 a single pool (K Pool, +32.1k USD) earns more than the entire net profit (+20.6k USD), i.e. the median held pool detracts.
3. Early-window scarcity. Few Lighter pools pass the gates before ~Nov 2025. The full-period vs holdout split already shows it: NB158 scores 0.72 Sharpe over the full period but 1.48 on the clean extension holdout.
4. Residual venue effects. Pool composition (discretionary perp-trader pools), data quality/discrete NAV jumps feeding the tail-driven result. Probably second-order given the population stats above.

How to test each driver

• A. Curation ablation (cheapest, most decisive): rerun the Hyperliquid NB158 baseline with top_n=None — the full eligible Hypercore population instead of the top-120-by-1Y-CAGR. If Hyperliquid collapses toward Lighter's numbers, the gap is mostly curation, not venue quality.
• B. Matched-breadth test: run the Hyperliquid baseline on ~20 random 34-vault subsamples of its universe; compare the Sharpe distribution against Lighter's 0.72. Prices the breadth penalty directly.
• C. Strategy-free equal-weight buy-and-hold index per venue over the identical backtest window from the daily share-price data: median vault return, % positive, pooled skew/kurtosis. Ground truth for "are the assets weaker over this exact window".
• D. Signal efficacy (rank IC): Spearman correlation between the age-ramp/Sharpe ranking at t and forward 30-day vault returns, per venue — does the signal predict on Lighter at all, or is there no cross-sectional dispersion to harvest?
• E–F. Age-aligned event-time return curves and BTC/ETH beta regressions if A–D leave a residual gap (BTC fell ~44% over the window).

One more piece of existing evidence: the Hyperliquid trials survived outlier-neutralisation (NB158 stayed strongly positive after removing its best day), while on Lighter all 31 methods go negative — consistent with the breadth story: Hyperliquid had enough winners to diversify the tail, Lighter has ~3–5 pools carrying everything.

Suggested next step: a single 03-research- notebook implementing A + B (both are reruns of existing machinery with a changed universe argument), with C/D added in the same notebook since the universe data is already loaded there.

🤖 Generated with Claude Code

[miohtama]

How to make a Lighter allocation strategy profitable

Refocusing from venue parity to strategy design: given 34 majority-losing pools, breadth-style allocation (hold ~20 names) is structurally unprofitable here — the trials confirm every diversified method is tail-driven. The design has to become selective and defensive. Grounding facts from the metrics feed and the NB01/NB02 runs:

• LLP (the protocol market-making pool) is the hurdle. 1.5y of history, 0% operator fee, net CAGR: 19.1% (1Y), 8.8% (3M), 4.7% (1M). Our NB01 strategy did 22.12% CAGR with a -26.65% drawdown — it barely beats passive LLP on return and almost certainly loses after risk. A Lighter strategy earns its complexity only if it beats LLP-plus-cash on risk-adjusted terms.
• The left tail is structural, and visible in advance. 7 of 34 eligible pools have lost ≥50% lifetime, and most are declared levered directional bets — "LINK 2x long" (-100%), "ETH 3x long" (-97%), "BTC 3x long" (-84%). These are leveraged market beta wearing a vault costume; holding them in a vault-of-vaults during a -44% BTC window does exactly what the label says.
• Exits are actually executable on Lighter. Pool withdrawal cooldown is 5 minutes (LIGHTER_POOL_LOCKUP), versus multi-day epochs elsewhere. Stop-loss discipline that is aspirational on other venues is real here.

Proposed design, mapped to tools we already have

1. LLP core + quality satellite. Make LLP the default allocation and park all un-deployed capital there instead of cash — the existing YieldManager queue-vault machinery (built for the phase-aware strategies) does exactly this. Satellites: 5–8 concentrated picks, per-name cap raised to ~25–30%.
2. Structural exclusion before statistical selection. Classify and exclude single-asset levered directional pools (name/description-based, curated into EXCLUDED_VAULTS). This removes most historical blow-ups with zero look-ahead — the leverage is in the pool's own label.
3. Hard quality gate for satellites. The NB02 trials tell us which tools work here: PSR confidence-scaling (NB54 — best full-period CAGR, 41.8%), trend overlay on the age ramp (NB101 — best holdout Sharpe, 1.49), and dynamic re-ranking with willingness to hold back (NB29 — best full Sharpe 1.18 at only -8.7% drawdown). Combine: PSR floor + minimum track record + trend quality, sized by confidence.
4. Cut the left tail mechanically. Per-pool drawdown-from-high-water-mark exit plus quarantine (the curator.py QUARANTINE_PERIODS machinery already exists), with no minimum hold on losers — keep slow-in (age ramp) but add fast-out. The 5-minute cooldown makes this honest in live trading.
5. Regime scaling. The BTC/ETH benchmark pairs are already loaded in the universe; most Lighter pools are directional, so scale satellite gross exposure down when the market regime is negative and let the LLP core carry the book.
6. Small data addition for a Lighter-native edge. Operator skin-in-the-game (pool_info.operator_shares / total_shares) and net flows exist in the Lighter public API but are not yet in our metrics feed (leader_fraction, netflow are unpopulated for chain 9998). A small eth_defi pipeline extension would unlock "operator has meaningful stake, low performance fee" as a forward selection prior — performance fees are already in the feed (0–30% spread across pools).

Expectations

A realistic target on today's universe is LLP-beating rather than Hyperliquid-matching: Sharpe ≥ 1.5, drawdown under ~10%, CAGR in the high teens to twenties — scaling up as the venue's pool count grows. Until then Lighter is best treated as a sleeve inside the existing cross-chain machinery rather than a standalone product.

Suggested notebook sequence

• 03-research-pool-quality-panel — classify pools (structural beta vs discretionary), LLP standalone baseline with drawdown profile, simulate the drawdown-exit + quarantine rule ex-post, and signal IC for the gate candidates.
• 04-backtest-llp-core-quality-satellite — LLP core + 5–8 pick satellite with PSR/trend gate, drawdown exits, regime scaling; benchmark against 100% LLP.
• 05-opt — optimise gate thresholds and stop levels against the clean extension holdout only.

🤖 Generated with Claude Code