Strategies

launch-sniper

src/strategies/launch-sniper.ts · quote token: WETH

Edge hypothesis: the first minutes after a NOXA instant-listing are the most information-rich and most volatile window a memecoin ever has. A disciplined buyer who only touches round-trippable, non-deployer-heavy launches, and exits mechanically instead of falling in love, harvests a slice of that opening volatility. The edge is speed + discipline, not prediction.

Entry filters (all must clear)

• Liquid route exists — a real QuoterV2 quote must resolve; un-graduated Odyssey bonding-curve tokens have no Uniswap pool yet and are skipped.
• Round-trip retention — buy then immediately quote a sell-back; a honeypot or thin pool that loses more than maxRoundTripLossPct (default 35%) of value is rejected.
• Deployer concentration — the launching wallet must hold under maxDeployerPct (default 15%) of total supply.
• Freshness — only launches seen within maxLaunchAgeSeconds (default 5 min) are considered.

Exits (mechanical, first match wins)

• Take-profit: takeProfitPct (default +60%)
• Stop-loss: stopLossPct (default −35%)
• Time exit: maxHoldSeconds (default 30 min), regardless of PnL

Failure modes

• Most new launches trend to zero — the stop loss fires often; profitability depends on winners covering losers.
• Honeypots can enable a sell tax after entry — the entry-time round-trip check cannot see that.
• Odyssey bonding-curve tokens are skipped pre-graduation — this is effectively a NOXA sniper.
• Paper fills use the QuoterV2 mid; live, faster bots win the best fills and you eat worse slippage.

momentum

src/strategies/momentum.ts · quote token: USDG

Edge hypothesis: a token that has already graduated to a locked Uniswap v3 pool survived the highest-mortality phase. A subsequent breakout — price up sharply over a short lookback — reflects real incoming demand rather than launch-day noise, and trend- following it with a trailing stop captures the middle of the move while giving back only the tail.

Entry

Price gain over lookbackSamples ticks (default 6) exceeds breakoutPct (default 15%) → buy entryUsdg (default $10) worth.

Exits

• Trailing stop: trailingStopPct (default 20%) off the post-entry peak
• Time exit: maxHoldSeconds (default 60 min)

Failure modes

• Breakouts on illiquid tokens are trivially fakeable by one wallet round-tripping the pool — this is a price-only signal with no depth check.
• Trailing stops whipsaw in choppy markets — expect a string of small losses between real trends.
• Discovery scans a bounded block lookback; a breakout on a token outside that window is missed until the next pass.
• No edge in a flat or declining market — this is a trend-only strategy by design.

premium-watch

src/strategies/premium-watch.ts · quote token: USDG · alerts-only by default

Edge hypothesis: Stock Token pools are much thinner than the underlying equity market, so DEX price can drift from the Chainlink oracle when flow is one-sided. If the drift is a liquidity artifact, buying the discount and later selling back at fair captures the reversion as the pool re-equilibrates.

Signal

Spread = (dexPrice - oracle) / oracle, in bps. Alerts when |spread| >= alertThresholdBps (default 50bps). Trades a discount (DEX cheap) when spread <= -tradeThresholdBps (default 150bps) — only in eligible+opted-in configuration.

Exits

• Convergence: spread reverts within exitThresholdBps (default 20bps)
• Time exit: maxHoldSeconds (default 2h)

Failure modes

• A premium/discount can be correct — DEX price may reflect information the ≤24h Chainlink heartbeat hasn't caught up to yet; fading it then loses on purpose.
• Stock Token pools are thin; the probe price doesn't reflect real trade-size slippage, and the trade itself moves the spread being arbitraged.
• No shorting primitive exists — only discounts are ever tradeable; premiums are alert-only by construction, not by choice.
• Chainlink staleness (weekend gap, feed outage) can masquerade as a discount when the oracle, not the pool, is stale.

llm-strategist optional, bring your own key

src/strategies/llm-strategist.ts · quote token: WETH · disabled unless configured

Edge hypothesis: launch-sniper's filters are fixed cutoffs applied independently — a launch either clears every threshold or it doesn't. An LLM can weigh the same real on-chain signals (round-trip retention, deployer concentration) holistically instead, and may make better marginal calls on launches that are close to the line. It sees nothing launch-sniper can't — same inputs, a different judgment function.

Signal

For each new launch, the strategy computes the same buy→sell round-trip retention and deployer-supply-concentration numbers launch-sniper uses, sends them (numbers only — no free-text on-chain fields, to limit prompt-injection surface) to the configured LLM with a strict JSON-only verdict contract, and parses the response with a tolerant extractor that throws on anything malformed.

Entry

Trades only when the verdict is {"buy": true} and confidence >= HOOD_LLM_MIN_CONFIDENCE (default 0.6). Everything else — a buy:false verdict, low confidence, a malformed response, or a provider timeout — is logged to the decision journal as an alert, never a trade.

Exits

• Take-profit / stop-loss / time-exit — identical bands to launch-sniper

Failure modes

• LLM confidence is not calibrated probability — a confidently wrong verdict trades exactly like a confidently right one.
• On-chain data reaching the prompt is attacker-controlled; only numeric signals are included today specifically to limit prompt-injection surface.
• LLM round-trip latency means this strategy reacts to a launch after launch-sniper already has (or hasn't).
• Every judged launch costs a real LLM API call regardless of verdict — a launch storm inflates spend with nothing to show for it.
• One configured provider, no fallback chain — a provider outage means zero coverage until it recovers.