88 lines
6.6 KiB
Markdown
88 lines
6.6 KiB
Markdown
|
# Could Jev Be Adapted to Outsmart the Book? (OB‑state quoting, fees & fill)
|
|||
|
|
|
|||
|
|
**Read‑only note. No data/CH/HZ touched.** Captures the reusable mapping of jev's
|
|||
|
|
`buildState` onto FLIGHT's fill‑prediction lattice (SRC‑01) + the EV gate, so a future
|
|||
|
|
session does not re‑derive it.
|
|||
|
|
|
|||
|
|
## TL;DR
|
|||
|
|
- jev's *placement* is naive: fixed `quoteInsideTicks=1`, fixed `tradeSizeMon=200`, one
|
|||
|
|
post‑only order **every block regardless of book state**. It earns the tick when a taker
|
|||
|
|
crosses it and loses the tick‑gas + opportunity cost when it does not.
|
|||
|
|
- "Outsmart the book" = predict fill probability from book **motion** (not the snapshot),
|
|||
|
|
gate `rest vs take` by the EV break‑even, size by the impact‑size curve, and **modify‑to‑the‑touch
|
|||
|
|
(fee‑free) instead of cancel+place**. All of that already exists in FLIGHT.
|
|||
|
|
- jev's `buildState` already assembles the *right inputs*; the missing piece is a
|
|||
|
|
**learned fill gate** over them. A general structured‑output LLM is the **wrong** tool
|
|||
|
|
— SRC‑01 shows a narrow learned factor lattice (the activity family + EWMA eigen)
|
|||
|
|
**beats** HD and a general LLM, and is far cheaper/faster. jev's value here is the
|
|||
|
|
clean state assembly + serialized cadence, not its oracle.
|
|||
|
|
|
|||
|
|
## 1. jev's book state — what it already sees (`model.ts` buildState)
|
|||
|
|
`bookImbalance`, `spreadBps`, `depthBps{10/25/50}`, top‑5 levels, `returnsBps{1/5/20/100}`
|
|||
|
|
(from a 400‑block `mids` window), `recentMids` (every 5th block over `horizonBlocks`),
|
|||
|
|
`taker` flow summary (`cvdMon`, `vwap`, `lastSide`). **The state is OB‑state rich.**
|
|||
|
|
The placement policy just ignores almost all of it.
|
|||
|
|
|
|||
|
|
## 2. SRC‑01 (FLIGHT's fill lattice) — what "outsmart" actually means
|
|||
|
|
Source: `EXEC_SOA_SUPERSESSION_AUDIT_20260808.md` §SRC‑01. Corpus 472 factors / 356,282 rows
|
|||
|
|
/ 8 assets / base fill rate 0.453 / metric = OOS within‑asset quintile lift (pp).
|
|||
|
|
|
|||
|
|
**The activity family dominates everything ("a book that is moving fills you"):**
|
|||
|
|
`n_chg_60 +18.54`, `q_path_60 +18.21`, `q_path_30 +17.10`, `n_chg_30 +16.67`,
|
|||
|
|
`fs_imbalance +16.42`, `q_range_60 +15.81`, `q_range_300 +14.98`, `q_range_15 +14.06`,
|
|||
|
|
`q_range_30 +13.63`. **Nothing in the OB snapshot comes close.**
|
|||
|
|
|
|||
|
|
- EWMA bank: `eig_ew_w750_lambda_max +11.4 → +12.5` at every span (10…1000) — robust, not an artifact.
|
|||
|
|
- `|vel_div|` (the NG7 incumbent) = **+5.02** — beaten 3.7× by the activity family, 2.5× by the EWMA eigen family. Velocity‑only underperforms.
|
|||
|
|
- **Impact‑size curve peaks at 200–2000 notional, collapses by 50,000.** Peak brackets **BLUE's actual $16,290 notional** — i.e. a resting offer *just* large enough to be reached by a ~30 s touch.
|
|||
|
|
- HD at 39 factors = **+20.73** vs best single **+19.17** — *barely* ahead; full 472 killed by memory; verdict OPEN. → a big general model does not pay off here.
|
|||
|
|
- New/odd: `oi_usd −3.77` (−OI ⇒ more pickup); **fewer book levels ⇒ more pickup** (thin‑book). Sporadic winner `eig_regime_signal +4.72` (z=+15.0, n=32,023) — well‑powered.
|
|||
|
|
|
|||
|
|
### jev state → SRC‑01 factor mapping
|
|||
|
|
| jev `buildState` | SRC‑01 factor(s) | "outsmart" lever |
|
|||
|
|
|---|---|---|
|
|||
|
|
| `returnsBps{1/5/20/100}`, `mids`(400), `recentMids` | `n_chg_60/30`, `q_path_60/30`, `q_range_*` | REST only when book is *moving in your direction* (p_fill high) |
|
|||
|
|
| `bookImbalance`, `depthBps{10/25/50}` | `fs_imbalance`, thin‑book effect | pick side where imbalance + thin depth ⇒ higher pickup |
|
|||
|
|
| (computed) book‑return eigen (EWMA bank) | `eig_ew_w750_lambda_max` | **highest‑value add** — jev doesn't compute this today |
|
|||
|
|
| `tradeSizeMon=200` (fixed) | impact‑size curve | SIZE to the curve peak (200–2000 notional); don't sit deep past reachability |
|
|||
|
|
| `quoteInsideTicks=1` (fixed) | (none — snapshot‑blind) | price inside touch *dynamically* by p_fill, not by fixed 1 tick |
|
|||
|
|
| n/a | `oi_usd` | avoid high‑OI windows (reduces pickup) |
|
|||
|
|
|
|||
|
|
## 3. The EV gate that decides rest‑vs‑take (the "avoid fees" hinge)
|
|||
|
|
`EV_rest − EV_take = p_fill × (Δ + offset) − (1 − p_fill) × s`, break‑even `p* = s / (Δ + s)`.
|
|||
|
|
`Δ = taker − maker = 4.50 − 1.50 = 3.00 bp` (measured, HL mainnet). `s` = realised cost of a
|
|||
|
|
maker miss (delay slippage + adverse drift).
|
|||
|
|
|
|||
|
|
- **Rest (maker, 1.5 bp rebate / capture the tick) only when `p_fill > p*`.** Below `p*`,
|
|||
|
|
you pay the ~300 ms block gas/wait and likely sit unfilled — that is the fee/opportunity cost jev
|
|||
|
|
currently burns every block unconditionally.
|
|||
|
|
- jev's replace‑every‑block **cancel+place** is *two* signed actions per cycle on Kuru; on
|
|||
|
|
HL that erodes the rate budget (`HlVenueAdapter` 1 req / $1 USDC traded — the #59 ledger).
|
|||
|
|
The fee‑free path to "outsmart + avoid fees" is **atomic modify‑to‑the‑touch**
|
|||
|
|
(FLIGHT `EntryReprice`, same cloid, ≤1/bar) — which jev does NOT have; jev cancel+places.
|
|||
|
|
|
|||
|
|
## 4. Why a structured‑output LLM is the wrong tool here
|
|||
|
|
- The signal is **exactly computable**: book‑change counts, path/range, imbalance, EWMA eigen —
|
|||
|
|
all cheap floats from the depth stream. The study proves a narrow learned lattice
|
|||
|
|
(activity family) **beats** HD and would beat a general LLM, at milliseconds and zero gas.
|
|||
|
|
- An LLM per‑cadence call is 50–200 ms of **added** latency against a venue already
|
|||
|
|
budget‑throttled (1 req/$1 USDC) and single‑writer‑lanned (ASEx P0‑P4). It is drag, not thrust.
|
|||
|
|
- jev's structured output is rightly confined to the **direction call** (where BLUE already
|
|||
|
|
owns the intent lane) — keep it there. The book‑outsmarting should be a **deterministic
|
|||
|
|
learned fill gate** over jev‑style state.
|
|||
|
|
|
|||
|
|
## 5. Adaptation recipe (jev's cadence + FLIGHT's gate)
|
|||
|
|
1. Keep jev's `buildState` (it is already OB‑state rich) + the serialized post‑only cadence.
|
|||
|
|
2. **Add** the activity‑family + EWMA‑eigen features (notably `eig_ew_w750_lambda_max`, today absent).
|
|||
|
|
3. Gate rest/take via the EV `p*` (`p_fill > p*` → rest; else take or skip) instead of
|
|||
|
|
"always rest 1 tick inside."
|
|||
|
|
4. Size by the impact curve (200–2000 notional peak; BLUE at $16.3k is the upper end of reachability).
|
|||
|
|
5. **Replace cancel+place every block with atomic modify‑to‑the‑touch** (FLIGHT `EntryReprice`,
|
|||
|
|
`offset_ticks=1`, `post_only=True`) to spare the HL rate budget — this alone is a big fee avoider.
|
|||
|
|
6. Reuse FLIGHT's `ExecControl.max_spread_bps` gate (exec_router `plan_entry`) — no point resting when spread > 5 bp.
|
|||
|
|
|
|||
|
|
**Net:** yes, jev‑style OB‑state quoting can outsmart the book; FLIGHT already built the
|
|||
|
|
outsmarting (SRC‑01 + EV gate + atomic modify). The adaptation is grafting jev's clean
|
|||
|
|
buildState + post‑only cadence onto that deterministic learned gate — **not** adding an LLM to the
|
|||
|
|
placement path.
|