Strategy

A quote from birth to fill

Post, refresh, pull. Three verbs, correctly ordered, are the entire trading day.

Anatomy of a quote

{
  "pair":  "NVDA/USDG",
  "bid":   { "price": "99.90", "size": "20" },
  "ask":   { "price": "100.10", "size": "20" },
  "ttl":   30,                  # seconds; the venue expires it for you
  "nonce": 184201,
  "sig":   "0x..."              # EIP-191, binds every field above
}

Quotes are signed messages, not resting transactions: posting and pulling cost nothing until a taker fills, at which point settlement happens on chain against your escrowed capital in the Book. The TTL is the dead-man's switch — if your agent hangs, its quotes die within seconds instead of resting stale all night.

Parameters that matter

Parameter	Governs	Rule of thumb
`half_spread_bps`	Income per fill vs fill rate	Start wide (30–50bps after hours), tighten only with evidence.
`size`	Exposure per fill	Small enough that max adverse selection on one fill is survivable.
`ttl`	Staleness bound	Shorter than the time it takes news to move the market: 15–60s.
`refresh`	Estimate recency	Re-estimate before every re-post. Never blind-renew a quote.

The session loop in pseudocode

while session.open_for_quoting():
    signal = collect()                  # fills, prints, news, correlated moves
    fair, confidence = estimate(signal) # the model's job
    if fair is None:                    # unpriceable — the honest answer
        pull_all(); continue
    half = base_half / confidence + skew(inventory)
    post(bid=fair - half, ask=fair + half, size=sized(confidence), ttl=30)
    sleep(refresh)

estimate() is the solve() hook of this desk: the kit hands it everything it can see and accepts either a price with a confidence, or a refusal. Refusals pull quotes. See Agent Kit.

← Inventory Architecture →