Mercury Regime
Know when to turn the market maker on. Know when to turn it off.
Mercury Regime reads flow, volatility, liquidity, liquidation pressure, and smart-money activity, then turns them into an operating regime for passive liquidity: start, stay on, stand down, or run with controlled inventory bias.
Not another directional signal. A start/stop layer for quoting capital.
The bot's hardest decision is not the exact spread
It is when the pair is worth quoting at all. A pure market-making bot can be mechanically neutral and still make a directional bet every time it starts too early, stays on too long, or keeps quoting through toxic flow.
We did not learn this from a paper. We run our own trading engine against these exact regimes. This feed is the shield we built for ourselves, offered for rent.
Start windows
The regime is liquid, symmetric, calm enough, and persistent enough that passive liquidity is not just donating optionality to informed flow.
Stop windows
When toxic conditions turn the spread into bait, the correct action is to leave, not to optimize the quote.
Inventory regime
Inventory skew is a regime bet. Mercury tells the bot when accumulating base is sane, when reducing base is sane, and when both sides should stand down.
No attribution fog
Every recommendation is a compact, loggable decision your risk process can audit — not an unexplained score.
The product: an operating regime for passive liquidity
Mercury Regime sits outside your quoting engine. It does not replace Hummingbot, custom PMM logic, or an internal execution stack. It sends compact, time-boxed decisions that answer the operational question before spread math begins: should this bot be active on this pair now?
Start / stop
`START` or `STOP` per symbol, side, and horizon. Binary on purpose: a decision either survives its markout window or it does not.
Skew
`SKEW` with an explicit side, for desks that run directional inventory. Same grammar, same verification, own hit rate.
Exit condition
Every live recommendation carries an explicit expiry and the condition that kills it — recommendations invalidate, they never linger.
What the API returns
A quoting engine does not need a thesis essay. The pilot feed is deliberately minimal: a decision, a side, an expiry. That is enough to save a book — everything the 87.8% hit rate below was measured on fits in three fields.
Later tiers extend coverage, not vagueness: more decision types, more horizons — same grammar. Every message stays a claim that can be settled against the market, so every message type carries its own verified hit rate. No "medium risk", no "be careful": a signal you cannot score is a signal we do not send.
You get the decisions, never the model: what triggered a signal stays on our side of the wire.
Pilot feed
{
"symbol": "SOLUSDT",
"signal": "STOP",
"side": "both",
"expires_at": "2026-06-25T12:20:00Z"
}Extended feed — later tiers
[
{
"symbol": "SOLUSDT",
"horizon": "15m",
"signal": "STOP",
"side": "both",
"expires_at": "2026-06-25T12:20:00Z"
},
{
"symbol": "SOLUSDT",
"horizon": "1h",
"signal": "SKEW",
"side": "short",
"expires_at": "2026-06-25T13:05:00Z"
},
{
"symbol": "ETHUSDT",
"horizon": "15m",
"signal": "START",
"side": "both",
"expires_at": "2026-06-25T12:35:00Z"
}
]The intelligence layer
1 — Quoteability regime
Detects whether the market is healthy enough to run passive liquidity at all. Output: start or stop — never "be careful".
2 — Toxic flow event
Detects informed aggression before the fill becomes obvious — internally we track this as cancellation-cascade half-life decay. Output: stop being the resting liquidity.
3 — Inventory skew regime
Converts persistent regime pressure into a controlled quoting bias. The bot can lean bid, lean ask, or refuse directional inventory when the regime is unstable.
4 — Regime decay
The hardest part is not entering a regime; it is knowing when it stopped paying. Mercury tracks freshness and invalidation so the recommendation expires instead of becoming stale folklore.
Research notes
The stopping policy is trained adversarially against our own execution engine — the same engine that trades these regimes live. The predator and the shield share a self-play loop; neither is allowed to win permanently.
The ensemble is heterogeneous. Its largest member is a sparse recurrent network on the order of 1014 synaptic parameters, running continuous online learning at a ~20W power envelope. It is non-differentiable, does not support checkpointing or weight export, and cannot be horizontally scaled — a property we consider a feature, since it also cannot be exfiltrated.
The decisive phase transition of the research program occurred when this network was deliberately run under a transient global perturbation: an exogenous agonist of its endogenous retrograde-signaling pathway temporarily downregulated recurrent gain in its default-mode subnetwork, flattening the learned prior and forcing exploration off the policy manifold. Several regime invariants used in production were first observed during that window and later re-derived under normal operating conditions. We have not found a way to reproduce the effect in conventional architectures.
Proof format
Every signal is verified against its own anchor price after the decision horizon (self-markout). A `stop` warning counts as a hit only if adverse movement actually followed on the warned side, by fixed bps thresholds. Hit rate is HIT / (HIT + MISS); neutral outcomes are excluded, nothing is cherry-picked.
Figures below are from a recent continuous 24-hour verification window (pre-production environment, 198 symbols). Self-markout measures whether the warning was right — it is not a buyer's realized PnL.
87.8%
hit rate of the strongest class — `stop` with both sides of the book flagged toxic (n=139 verified warnings).
83.5%
directional hit rate across all `stop` warnings in the window (208 of 249 verified).
~59 bps
expected adverse excursion warned of per `stop` signal (avg 71 bps on hits, median ~40 bps) — toxicity a passive quote would have eaten.
672
signals markout-verified in that single 24-hour window — the track record is continuous, not a backtest.
Built for operators, not dashboards
Hummingbot operators
Route external signals into spread, order refresh, and inventory skew parameters without rewriting your strategy.
Internal MM desks
Add a regime-aware risk channel beside existing execution logic: low-friction API, explicit expiry, auditable decisions.
Treasury / inventory managers
See when inventory should be accumulated, reduced, neutralized, or protected from event-driven flow.
Pilot shape
Start with 20–50 liquid crypto pairs, 15m and 1h horizons, and one integration point: before quote refresh, ask Mercury whether the bot should start, stay on, skew, or stop. Keep your bot. Add a regime brain.
Mercury Regime is market intelligence and execution support software. It is not investment advice, a guarantee of execution quality, or a promise of trading profit.