Quantitative Algo Engines (furlpay algo)
The research-driven analysis surface — every engine implements a formula from the HFT V2 research report, runs offline over data you supply, and fails closed on malformed input.
Honesty scope
Schedules are plans, verdicts are arithmetic. Nothing here places, signs or broadcasts an order, and no live edge is ever claimed. When the WebSocket engine ships, it feeds these same pure engines.
algo quotes — GLFT closed-form quotes
The Guéant–Lehalle–Fernandez-Tapia model — the deployable quoting frontier per the 2026 literature (hftbacktest ships a full implementation), adding the ½γσ²τ volatility term that the Avellaneda-Stoikov variant omits:
δa + δb = γσ²τ + (2/γ)·ln(1 + γ/k) (total spread)
r = s − q·γσ²τ (inventory-skewed reservation)
p(δ) = e^(−kδ) (GLFT fill probability)furlpay algo quotes --mid 6482150 --gamma 0.1 --sigma 0.02 --kappa 1.5 [--inventory q] [--time-left 1]algo schedule — micro-TWAP / POV execution planning
Poisson-jittered time slices (deterministic for a pinned --seed), sizes reconciling exactly to the parent amount in BigInt — no minor unit minted or lost. POV mode caps participation at --cap (default 3.5%) over a supplied per-minute volume profile and fails closed without one — the tool refuses to guess volume.
furlpay algo schedule --amount 1000000000 --duration 900 --slices 12 --seed 42
furlpay algo schedule --amount 5000000 --duration 3600 --mode pov --volume profile.json --cap 0.035An incomplete POV plan (observed volume < parent amount at the cap) reports the remainder — it never auto-chases the tail, because chasing the tail is how POV orders become impact events.
algo report — Implementation Shortfall scorecard
IS is measured against your arrival price — the only honest scorecard, since it's the decision price, not a mid the executor can pick. Exact BigInt accounting: Σ size·(price − arrival), VWAP with the division remainder surfaced, fees in minor units and bps.
furlpay algo report --arrival 10000 --fills fills.json # [{ priceMinor, sizeMinor, feeMinor? }]algo breakeven — maker ship/no-ship arithmetic
The §7.1 objective function as a decision tool:
E[PnL/hour] = λ·(δ − a − c) − ½γσ²q² (bps)λ is your observed fill rate at that spread — the fill-probability adjustment (arXiv:2502.18625) enters through λ: widening δ lowers λ, and the tool refuses to pretend otherwise.
furlpay algo breakeven --half-spread 2 --toxicity 1 --fee 0.5 --fills 10 --inventory 1 --gamma 0.1 --sigma 0.02algo queue — fluid queue-position estimate
Transparent fluid model: the queue ahead drains at (trade rate + cancel rate). Estimates seconds-to-front and fill probability within your horizon. Useful for ALO quoting sanity — if the queue can't clear in your horizon, the quote is dead inventory. (Approximation, not the full Moallemi queue-value model.)
furlpay algo queue --ahead 100000 --trade-rate 150 --cancel-rate 50 --horizon 60algo funding — persistence z-score + honest carry
Funding entry needs both a z-score above threshold (default 2.0) over your supplied ≥8-sample per-8h history and operator OI confirmation. A single high print is not a signal. Annualized gross uses f̄·3·365; the verdict reminds you that crowding compresses funding, both legs pay fees, and basis drift forces rebalances.
furlpay algo funding --history funding.json --z 2.0 --fees 8algo cascade — liquidation-cascade regime flag
Composite of the three cascade signatures: funding spike (≥5 bps/8h), ask-wall OBI (≤ −0.65), toxic VPIN (≥ 0.75). Score ≥ 2 → CASCADE-LIKELY → reduce maker exposure.
There is no sniping edge
Per the V2 report §5 correction: Hyperliquid routes liquidations through the book without discount; failed liquidations go to the HLP backstop (which you can participate in live via earn hlp). The cascade tool is analysis only — every response carries this correction.
furlpay algo cascade --funding 0.0006 --obi -0.7 --vpin 0.8MCP exposure (read-only)
AI agents get the same math through three read-only tools: algo_glft_quotes, algo_breakeven, algo_cascade — annotated readOnlyHint, offline-labeled in every response. See MCP Agents.