Superior Skills

“Open agent skills and tool schemas for Superior Trade — build, backtest, and deploy trading strategies on Hyperliquid.” MIT-licensed, 225★, last pushed 2026-07-27. A library of Agent Skills-standard folders — one SKILL.md each — covering strategy templates, reusable primitives, and venue integrations for Hyperliquid, Aerodrome, Lighter and Polymarket.

The code is open; the platform behind it is not. Every skill that does anything needs a SUPERIOR_TRADE_API_KEY against api.superior.trade. See agent-deployed-trading for what that arrangement is.

The two “validated” strategies

The README distinguishes validated strategies (“with backtest evidence”) from template strategies (“starting points”). Two are validated, both on a single 162-day window, 2025-11-20 → 2026-05-01:

The first row is not evidence of anything. Six trades over 162 days, a 100% win rate and a 0% maximum drawdown is the signature this corpus was warned about in the same week: ernest-chan built a whole book around “simple and linear strategies, as an antidote to the overfitting and data-snooping biases” (chan-algorithmic-trading). A sample of six cannot distinguish a strategy from a filter that happened to fire on six good days, and a 0% drawdown over five months is what a rule that trades almost never looks like, not what a robust rule looks like. The label “validated” is doing work the number cannot support.

The second row is a real result by comparison — 84 trades, four pairs, an honest 18.5% drawdown, and a win rate well short of perfect. It is still one window, one venue, one author’s backtester.

The pairing claim is the interesting part: run as separate sub-accounts, the two are said to be regime-complementary — the Donchian gate fires zero trades in the chop windows where the Bollinger reverter works, and underperforms in the trend windows the Donchian captures. That is a testable structural claim rather than a return figure, and it is the version of market-regime-analysis this spoke has been describing abstractly. Nothing here tests it out of sample.

The library

Eight strategy templates across five families the spoke already names on algorithmic-trading — trend-following (breakout, Donchian), mean-reversion (Bollinger fade), carry (funding-rate arbitrage, funding squeeze), arbitrage (spot-perp basis), scalping (RSI + volume thrust) — plus dca-weekly and grid-trading.

More useful are the primitives, which are the reusable parts factored out:

• regime-overlay — a triple-confirmation gate (EMA separation + ADX + N-bar return), pitched as what “turns fragile directional strategies into regime-robust ones.”
• dsl-exit-engine — three-phase exit: ROI ladder, hard stop, ratcheting trailing stop.
• fees-optimizations — maker (ALO) versus taker order-type decisioning, builder fees, parameter sweeps. Fee modelling as a first-class concern is rare in this corpus.
• trade-thesis — a structured pre-trade write-up: bull and bear cases, invalidation criteria, sizing rationale, required before any new strategy goes live.
• backtesting — window selection, walk-forward, parameter sweeps.
• intelligence — an opportunity scanner ranking pairs through a four-bucket funnel (squeeze fuel, stealth accumulation, coiled spring, basis flipping).

The stated workflow is draft → backtest → review → deploy, with “no live deployment without explicit user confirmation,” every parameter logged and every trade tracked. There is an atomic exit-all kill switch per venue.

Its risk disclosure, which is better than most

Quoted because this corpus’s sources usually skip it:

Trading involves risk. Backtests do not guarantee future performance. The validated strategies above showed positive returns on a single 162-day window; a strategy that worked then may not work in a different regime… Pair every deployment with the regime-overlay gate or equivalent — strategies without regime confirmation are demonstrably fragile.

It names its own window, declines to annualize, and says outright that ungated directional strategies are fragile. Against smolinski-h1-2026-aar‘s unaudited +33% or the founding bots’ self-reported figures, that is a step up in candour — which makes the six-trade “validated” table sit more awkwardly, not less.

Tier

T3 — vendor-published, first-party backtests run on the vendor’s own engine, with the strategies selected by the party selling access to them. No independent replication, no out-of-sample window, no Sharpe, no volatility, no trade log. The skills themselves are inspectable and MIT-licensed; the backtester, the data and the execution are behind an API key, so the numbers cannot be checked by anyone outside.

Not T4 only because the disclosure is specific and the losing conditions are stated. The one figure a reader should carry away is the sample size, not the return.

Related

agent-deployed-trading · backtesting · market-regime-analysis · mean-reversion · algorithmic-trading · chan-algorithmic-trading · strategy-optimization · event-driven-trading · synthesis