A spread, narrowly defined, is the distance between the best bid and the best offer at a given moment in a given instrument. That definition is also, for the institutional desks reading this, the least interesting thing about it. The interesting question is: what does that distance encode? Inventory risk. Adverse-selection probability. Latency to the next price update. The probability that the venue you would hedge against is, right now, showing your size to someone else.
This is the first in a five-part editorial series on how Drovix is built. It explains why a tighter institutional FX spread is not a marketing claim and not a subsidy, but the deterministic output of an architecture that takes those four costs seriously every microsecond of every trading day.
Across the series we cover: how Drovix prices (this post); how regulated retail brokers hedge their B-book through us; how our smart order router decides when to route flow externally; the latency budget that keeps all of it possible; and the pre-trade risk and audit framework that makes the whole stack auditable for compliance.
Three things every institutional FX quote must price
Strip away the surface and a two-sided quote is a wager on three independent random variables. Most engines try to wager on only one of them. The tight, fillable spreads that distinguish wholesale liquidity from retail come from systematically pricing all three.
1. The arrival probability of toxic order flow
Some counterparties carry information. Internalising their order means you have just been the slowest reader of public truth. A market maker that does not measure adverse selection — by counterparty, by symbol, by hour of day, by macro regime — eventually subsidises the informed and overcharges the rest. The result is a book that looks tight on a Wednesday afternoon and goes wide every CPI Wednesday morning, because the engine never learned which 0.2 % of takers cause 70 % of the realised toxicity.
Drovix maintains a continuously updated toxicity score per counterparty, calibrated on realised post-fill price moves. The score is one of the inputs the price-construction model uses to skew its outbound quote. The point is not to refuse flow — it is to price flow honestly.
2. The cost of the hedge that has not happened yet
Every time you stand a firm quote, you are implicitly underwriting the cost of unwinding the position if it is hit. A quote is only fair if that hedge is realistic, in this size, at this minute, on the venues you actually have credit at. A 0.1-pip headline spread is theatre if the offsetting venue widens to 2 pips before your hedge clears, or if your largest LP throws a 12 % reject rate the moment the print hits.
A genuinely institutional FX market maker prices the expected hedge cost — not the best-case one. That means a model that observes venue health continuously, not a config file that updates on quarterly reviews.
3. The optionality you have given away
A standing quote is a free option for the taker. The fair price of that option depends on instrument volatility, on the speed of the next state transition, and on how stale your own price is by the time it is hit. Most price engines under-price this option; the engineering fix is not to charge more, the fix is to be faster. Speed is the cheapest way to compress optionality, which is why a serious institutional engine treats latency as a product feature, not an operational concern.
How Drovix actually builds an institutional spread
Three architecture decisions, all boring, all unglamorous, that compound into a structural advantage over any single-LP arrangement or commodity ECN.
A. Aggregate at the source, not at the screen
Drovix maintains direct connectivity to tier-1 bank and non-bank liquidity providers and to selected lit venues. Their books are normalised and aggregated inside our matching engine — not stitched together in a UI layer, not re-priced through a customer-facing API. A taker on the institutional book is not seeing one LP's view of the world; they are seeing the joined distribution.
The practical effect is that depth survives the trip from raw FIX feed to executable quote. Aggregating at the screen typically loses the second and third levels of book; aggregating at the source preserves them, which is what makes 5–10 lot fills routine on majors that look like 1-lot books elsewhere.
B. Skew the quote with an ONNX model that lives next to the order book
Our price-construction model is trained off-line on years of order-book microstructure and exported as an ONNX runtime graph that runs on the same machine that emits quotes. Inference happens in tens of microseconds. That model continuously adjusts the skew of our two-sided quote based on observable features — order-book imbalance, recent micro-volatility, the counterparty's historical toxicity score, time since last price update — and it does so before a competitor with an off-machine inference layer has finished its first RPC.
Putting the model in-process matters more than the model's clever-ness. A perfect model that takes 3 milliseconds to evaluate is, in production, a worse model than an 85th-percentile model that takes 30 microseconds. The reason is that the market moved during those 2.97 milliseconds of network latency, and the perfect model's output is now valid for a state of the world that no longer exists.
C. Withdraw cleanly, never partially
A bad quote is worse than no quote. When the model's confidence collapses — typically just before scheduled releases, or in the seconds after an unscheduled headline — Drovix widens or withdraws synchronously across all our outbound prices. We do not leave a stale firm side hanging on one venue while a fresher side runs on another. That synchronicity is itself a function of the underlying transport layer, which we cover in Microseconds Matter.
Why this matters for retail brokers hedging at Drovix
A regulated retail brokerage that internalises end-customer flow is, by definition, taking the other side of a stochastic stream of orders. The economic question is not whether to hedge, but at what cost. A wide hedging spread eats the very alpha that makes the B-book worth keeping; a hedging counterparty that withdraws under stress turns a calm Monday into an exposed Wednesday.
Drovix prices that engagement explicitly as a wholesale relationship — known counterparty, known size, known cadence, principal-to-principal. We make our money on volume, transparently disclosed, and we are not in the business of trading against the retained internalised book that pays your operations team. That alignment is the precondition for a tight spread that survives stress. We elaborate the full operational pathway in B-Book to Wholesale.
Why it matters for hedge funds and proprietary trading firms
If you take liquidity professionally, the cost is the spread plus the impact of being read. A counterparty that cannot route beyond its own internal book is, structurally, going to leak your intent — and once your intent leaks, the price moves before your full size clears. Drovix is not designed as a terminal venue; when our internal price is not the best available, we route. The architecture of that routing is the subject of Routing Beyond the Inside Quote.
How we measure ourselves
Tight spread is a self-serving metric. The metrics we publish to our institutional counterparties on the portal are the ones that survive cross-examination by a quant team:
- Effective spread — the realised round-trip cost on filled orders, including any slippage from the displayed price.
- Fill rate at quoted price, in size — the percentage of orders that filled at the price we displayed, segmented by notional band.
- Time-to-fill p50 and p99 — both, because a tight average with a long tail is a marketing trick, not a product.
- Reject rate and reject taxonomy — credit, last look, technical, market — so the counterparty can argue with any of them.
- Price improvement frequency — when we beat our own quoted price, how often it happens, and the average bps of improvement.
Anyone can advertise an inside spread. The honest measure is what arrives in your TCA report on Friday afternoon.
What a fair spread is not
A fair institutional spread is not the tightest spread on a marketing page; it is the tightest spread you can actually fill in size, repeatably, at 02:00 on a Sunday evening Asia open and at 13:30 on a CPI Wednesday afternoon. Anyone can advertise zero — the institutional question is what arrives in TCA after the fact, and whether the engine telling you the number is the same engine that filled the order.
Tight spreads are an outcome, not a feature. The feature is the architecture that lets them stay tight when it would be easier to give up.
Where to read next
→ B-Book to Wholesale — a practical playbook for regulated retail brokers moving warehoused B-book risk into a principal hedge at Drovix.
→ Routing Beyond the Inside Quote — how the Drovix smart order router decides when to send flow outside, and to whom.
→ Microseconds Matter — the engineering deep-dive: C++ on pinned cores, Aeron transport, on-machine ONNX inference.
→ Risk Without Friction — pre-trade gates, multi-party approvals, and the immutable journal that makes the whole stack auditable.
Analyst Desk
Drovix Research Desk
Institutional Research
Drovix Research Desk publishes institutional-grade analysis covering macro events, cross-asset correlations, and execution insights for professional market participants.