research: Lighter leveraged-beta exclusion and LLP-core backtests NB03-NB05

[miohtama]

What

Continues the Lighter vault-of-vaults chain (NB01–NB02, #44) to its allocation conclusion. Three executed notebooks:

03-backtest-rebalancing-trials-no-leveraged-beta.ipynb

All 31 catalogued rebalancing methods rerun on the 27-pool universe with the 7 declared single-asset leveraged directional pools (BTC/ETH 3x long+short, SOL 3x short, HYPE 2x long, LINK 2x long) excluded via new curator blacklist rules, 5 positions, 33% concentration cap. Headline metrics improve sharply (best holdout Sharpe 1.49 → 3.38, NB29; the age-ramp waterfall family collapses to worst), but outlier-neutralising the top 5% of profitable days still turns every method negative — the tail-dependence is structural.

04-backtest-equal-vs-inverse-vol-baselines.ipynb

Ports the Hyperliquid survivorship-free NB26 baselines (composite selection; equal weight vs inverse-vol + 30-day-return gate). The damage gate halves the drawdown (−41% → −22%) and flips skew positive, but neither baseline is profitable on this universe over the window.

05-backtest-llp-core-healthy-satellites.ipynb

Implements the point-in-time redesign — LLP core plus hygiene-gated inverse-vol satellites (TVL ≥ $100k, < 50% stale marks, TVL ≥ 25% of running peak, ret30 damage gate), weekly cadence — against a 100% LLP benchmark. Two findings beyond the headline:

• A real framework hazard, found and fixed: with raw signals not summing to one, normalise_weights() redistributed the capped core's excess into the sole qualifying satellite, putting Guinea Pool at 3× its cap days before its collapse (−21% single day). Fix: satellite caps enforced in raw-signal construction, core absorbs unused budget.
• Fee/slippage verification: all Lighter pools have zero deposit/withdrawal/management fees, performance fees are internalised in the share price, and the backtest fills vault trades exactly at NAV with no cost (slippage_tolerance_pct is a live-execution guard only). The verdict is not a cost artefact. This is now codified as the CLAUDE.md "Vault backtests" rule (arriving via the research/allocation-predictive-features branch).

Final verdict: 100% LLP dominates — Sharpe 2.33, +18% CAGR, −4.8% max drawdown versus 1.16 / +15% / −12.5% for the best active construction. The strategic conclusion is in the first PR comment.

Dependencies

Needs the pending trade-executor build_lighter_vault_universe(exclude_leveraged_beta=True) blacklist-rule change (currently in the working tree; companion PR to follow) on top of merged trade-executor #1572 / trading-strategy #241 / web3-ethereum-defi #1348.

🤖 Generated with Claude Code

[miohtama]

Is there a point to portfolio construction on Lighter at all, given LLP's dominance?

Within Lighter itself, today: no — as a return-seeking exercise, multi-pool construction is not worth it. NB01–NB05 tested that hypothesis to destruction: cost-free satellites selected by validated point-in-time hygiene rules still subtracted ~3pp of CAGR and doubled the drawdown against simply holding LLP. With ~6 live satellites totalling $6M TVL ($1M deployable at the 20% pool cap), zero measurable selection IC, and LLP holding ~92% of all live TVL, there is no breadth to construct a portfolio from. The rational Lighter allocation is 100% LLP.

Three qualifications, which are where the value of this research arc sits:

1. The construction research is what earned the "just hold LLP" conclusion — and keeps it honest. It is now a validated, time-stamped result rather than an assumption. It is also unlikely to be permanent: LLP's ~19% 1Y net return comes from venue volumes and liquidation revenue spread over pool TVL, and per-dollar yield of protocol MM pools typically compresses as TVL grows (HLP's did). The hygiene indicators built here double as the monitoring that says when this conclusion expires.

2. The portfolio construction that matters is one level up — sizing LLP inside a multi-venue book. 100% LLP is a concentrated bet on one protocol's smart contract, one operator, one rollup, and a pool that is structurally short tail risk: it is the venue's market-maker and liquidation backstop, earning smooth daily premia and eating the violent days (skew -3.2, kurtosis 94, -5% worst day — a 2.33 Sharpe with that shape is partly compensation for rare large hits, possibly flattered by smoothed marks). Lighter-internal satellites cannot diversify any of that — they share the same venue and rollup. The diversification that works is cross-venue: LLP alongside HLP-style pools and the EVM vault universe in the existing cross-chain machinery, where the right question becomes "what venue-risk cap does an LLP sleeve get — 10–20% of the book?" rather than "which Lighter trader pools to pick".

3. The machinery is an option on the venue maturing. Lighter went from ~1 pool with ≥$7.5k TVL to ~22 over the backtest window. The Hyperliquid comparison is instructive: the same vault-of-vaults approach did beat the anchor pool there, because that universe has 350+ pools with genuine (if survivorship-tilted) winners. Portfolio construction pays where breadth exists; Lighter does not have breadth yet. The core/satellite allocator, hygiene rules, and damage gates are built, debugged (including the weight-normalisation hazard), and benchmarked — ready to switch on when it does.

Recommendation: fold Lighter into the cross-chain vault-of-vaults as an LLP sleeve with a venue-risk cap, keep the hygiene indicators running as monitoring, and shelve Lighter-internal multi-pool allocation behind explicit revisit triggers — e.g. ≥15 pools with ≥$500k live TVL and fresh marks, satellite deployable capacity ≥$5M at the 20% pool cap, or a measured positive selection IC over a fresh 6-month window. Until one of those fires, further Lighter-internal allocation research would be rearranging six deck chairs around one ocean liner.

🤖 Generated with Claude Code