Pricing every FX customer individually

A major Turkish bank had a pricing problem that did not look like one. Every retail FX customer was getting the same spread, posted each morning against the same market. Margin was steady, peer-comparable, defensible. The desk thought it was doing its job. One spread for everyone is the worst pricing strategy on both ends of the book, overpricing the customers who would walk on a tighter quote, underpricing the customers who would have transacted anyway.

The context matters. In Turkey, household savings sit substantially in USD, EUR, and gold. Retail FX is not a discretionary product. It is the way a meaningful share of the population stores wealth. The bank’s desk handled flow from millions of customers at both ends: heavy transactors moving large positions weekly, and ordinary customers buying a few hundred dollars on payday. One spread treated all of them as the same buyer.

The bank’s commercial leadership already knew the spread had elasticity buried in it; they had read the literature and run internal models that hinted at structure. The harder question sat above the pricing decision: where the bank wanted to position itself among its peers in retail FX, whether premium and slower, volume and tighter, or something segment-specific that one published spread could not express. What they could not find was a firm that would do both halves of the work with them — define the strategic position alongside the bank, and then ship a system that priced each customer differently at scale, on production infrastructure, validated against a live customer base.

By the time the proposal arrived, a price-sensitivity inference layer was already running against the bank’s own transaction history, pulled under a discovery agreement and tested on held-out months. No bank in any market had published per-customer FX pricing at this granularity; the doability risk had to live somewhere, and the natural place was on the proposing side, before the engagement opened. The proposal led with the output, not a description of it.

The architecture was four models in series rather than one end-to-end model: sensitivity, segmentation, per-quote spread, and a volume-versus-margin lever the bank could tune explicitly. The composition kept the system explainable, auditable, and editable at each layer.

The model could not answer every question. Dormant customers, new customers, and market regime shifts each fell outside what behavioural inference could cover. Those were business decisions the bank made in the same window: an attractive entry spread for new and dormant accounts, a conservative default during volatility shocks, named owners on each call.

Six months from kickoff to production. The pilot lift was asymmetric: heavy transactors produced a 22% lift in profit, ordinary customers, the long tail of payday FX buyers, produced 68%. Ordinary customers got tighter quotes than the published spread and transacted more often; the 68% lift came from volume, not from charging each customer more. Heavy transactors received quotes closer to what their behaviour already showed they would accept, where the published spread had been leaving margin.

Personalised pricing in financial services has to survive a fairness question. The four-model architecture answers it: every quote is traceable to the elasticity, segment, and market state that produced it, and defensible on those terms to the board, the regulator, or a customer who asks.

The system rolled out across the full base after the pilot validated and has been running there since. The engagement defined the governance frame alongside the build: override authority with the trading desk, retraining on a fixed cadence with sign-off, the volume-versus-margin lever held by the head of retail. The engagement closed when the system was operating and the governance had moved to the bank.

The second-order outcome matters as much as the profit lift. The bank moved from following peer-published spreads to setting its own price discipline. Pricing became an internal capability rather than a tracked market signal, the part of the engagement that is hardest to reverse.