First time I decided to head into e-trading was when I was working at Lehman Brothers on the e-procurement side. You could see the ecommerce teams starting to grow and the feelings on the trading desks started to change in how liquidity was managed. My choices in the e-trading space where always built on technology even though the mechanics of the trades between buyside and sellside where relationship based. Some in the e-trading space still argue relationships are key to liquidity…but are they?
In the noughties a dealer knew which corporate treasurer would trade at the London fix, which macro fund would show size into New York, and which hedge fund would pull liquidity the moment volatility picked up. Prices were formed as much by relationships and judgement as by supply and demand. Away from the trading side the actual apparatus of which trading technology was selected from an array of vendors was also often dictated by relationships between client and vendor.
As we enter 2026 I think you can argue that world has largely disappeared or is in the process of disappearing. FX liquidity is no longer negotiated; it is parametrised. It arrives as streams, curves, confidence scores and reject probabilities. What used to live in a dealer’s head now lives in configuration files and datasets.
This shift from voice to code is not just a story about automation. It is a story about how FX liquidity itself was redefined. From human discretion to machine-readable parameters and how that change reshaped trading behaviour, market structure, data economics and the regulatory environment.
The voice market: liquidity as judgement
In the voice-driven FX market, liquidity was inherently contextual.
A quoted price depended on:
- Who was asking
- How much they wanted to trade
- Why they were trading
- When they were likely to come back
A dealer could widen a price not because volatility had increased, but because they suspected the client had information. They could show size selectively, lean on internal flows, or warehouse risk based on experience rather than models.
Importantly, liquidity was elastic. It could be negotiated, delayed, reshaped, or withheld entirely. The concept of a “firm price” was flexible, and execution quality was inseparable from relationships. Data existed — but it was secondary. The primary signal was conversation either on the phone or on the Bloomberg/Reuters chat.
Electronic trading didn’t remove discretion — it encoded it
The first wave of electronic FX trading did not eliminate dealer judgement. It translated it. Single-dealer platforms (SDPs) allowed banks to stream prices to clients, but those prices were still shaped by:
- Client tiering
- Historical behaviour
- Internal risk limits
- Dealer intuition
What changed was the interface, not the decision-making. Liquidity became continuous rather than episodic, but it was still deeply relationship driven. The crucial shift came later, when multi-dealer platforms (MDPs) and algorithmic execution forced liquidity to become comparable. In 2000 FXConnect, FXAll, Currenex and 360T launched and more MDPs followed. Once prices from multiple banks appeared side by side, discretion had to be expressed in a way machines could process. That required parameters.
Parametrisation: the moment liquidity became machine-readable
In FX, this happened gradually but decisively. Dealers began to express liquidity through:
- Spread widths
- Skew adjustments
- Size tiers
- Timeouts
- Reject logic
- Last look thresholds
- Client tiering levels (Gold, Silver, Bronze, etc)
Each parameter encoded a piece of human decision-making:
- How much risk am I willing to take?
- How confident am I in this price?
- How toxic do I think this flow is?
- How fast do I want to respond?
Liquidity stopped being a conversation and became a function.
From the buy-side perspective, this was transformative. Instead of negotiating, traders could probe. They could send RFQs, stream requests, and child orders to infer liquidity conditions statistically rather than socially.
Execution algos then formalised the new language of liquidity
The rise of FX execution algorithms completed the transition.
Execution algos required liquidity to be:
- Observable
- Predictable
- Quantifiable
An algo cannot “sense the market” the way a human trader once did. It needs inputs. As a result, liquidity had to be broken down into measurable components:
- Fill probability
- Market impact
- Slippage distribution
- Latency sensitivity
This forced both sides of the market into a feedback loop. Banks tuned their parameters to protect against adverse selection. Buy-side firms measured those parameters implicitly by analysing outcomes. Over time, liquidity became something inferred from data outputs rather than shown explicitly through relationships. In this sense, execution algos did not just consume liquidity they contributed to reshaping it.
As liquidity became parametrised, buy-side firms adapted by becoming data businesses
Why “liquidity” in FX no longer means what it used to
In a parametrised market, liquidity is not depth at the top of book. It is a conditional probability.
A tight price is only meaningful if:
- It is firm
- It survives latency
- It fills at the expected size and price
- It does not disappear on an execution attempt
This is why FX liquidity often looks abundant until it isn’t. During normal conditions, parameterised liquidity performs well. Models are calibrated on stable regimes, reject rates are low, and spreads behave predictably.
Under stress, parameters flip:
- Size thresholds drop
- Last look windows tighten
- Skews widen asymmetrically
- Streams pause entirely
What disappears is not liquidity itself, but the assumptions embedded in the parameters. Is this when relationships matter or when the data aligns?
The buy-side response: measuring what cannot be seen
As liquidity became parametrised, buy-side firms adapted by becoming data businesses. The last 10 years have seen this increase rapidly across the buyside, although perhaps less so with corporate treasurers.
Buyside Execution desks began to capture and analyse:
- Quote-to-trade ratios
- Reject reasons
- Time-to-fill distributions
- Venue-specific performance
- LP behaviour by pair, size, and time of day
This data became was key to a change in mindset on the buyside. Over time, sophisticated buy-side firms stopped asking “where is the best price?” and started asking:
- Where is the most reliable liquidity?
- Which LPs behave consistently in stress?
- How does liquidity decay as size increases?
These questions can only be answered statistically — another sign that liquidity had become abstracted from human interaction.
Venue evolution: from execution to data infrastructure
The parametrisation of liquidity could also be put forward as another strategic reason exchanges acquired FX platforms. I have written on this in previous articles but deals such as CME and EBS, LSEG and Refinitiv, Deutsche Börse and 360T and BidFX and SGX. They were not simply about volume. They were about owning the rails on which parameterised liquidity flows.
Venues increasingly act as:
- Normalisers of heterogeneous liquidity
- Distributors of analytics
- Repositories of historical behaviour
In a world where liquidity is defined by parameters, the ability to standardise, measure, and replay those parameters becomes strategically critical. Execution is ephemeral. Data persists.
What was lost — and what was gained
The parametrisation of FX liquidity brought undeniable benefits:
- Lower transaction costs
- Greater transparency
- Scalability across regions and time zones
- Reduced reliance on individual dealers
But something was lost as well.
Human discretion once absorbed ambiguity. A dealer could choose to show liquidity despite uncertainty, based on judgement. Parameterised systems are less forgiving. When uncertainty rises, the default response is often to withdraw. This is why FX liquidity can feel binary for some clients: abundant until suddenly absent.
The next phase: adaptive parametrisation
The future of FX liquidity is never going to be a return to voice, nor a simple extension of current algos. It is adaptive parametrisation.
This includes:
- Dynamic skewing based on real-time flow toxicity
- Machine learning models for reject probability
- Venue selection that adapts intra-order
- Feedback loops that update parameters continuously
But even here, the core truth remains, that liquidity is still being expressed through parameters. The difference is that those parameters are now adjusted faster and with more data. The market has not become less human — it has become human judgement at scale, encoded in systems. Some of the eFX platforms have seen this and anyone dialled into looking behind the press releases can see that strategically some platforms are well in advance of others in terms of future proofing.
We haven’t even approached how AI will affect these MDPs and the trading/liquidity on them. It’s a subject for a larger piece but look at what some of the sharpest minds in the biz say about AI and its impact on trading firms operationally. See the remarks recently in the press from Citadel CTO Umesh Subramanian and Schonfelds Ryan Tolkin.
Conclusion: liquidity as a design choice
FX liquidity did not disappear. It was redesigned. What was once negotiated became calculated. What was once implicit became explicit. What was once personal became statistical. Understanding modern FX markets requires understanding this transformation. Not just how algos work, but how liquidity itself was turned into code and what that means when the assumptions behind that code are tested.
Readers can see more of John McGrath’s articles on his Substack page: johnjmcgrath.substack.com
