Skip to content

ADR-0015: Cross-rates divide through a pivot currency

ADR-0015: Cross-rates divide through a pivot currency

Section titled “ADR-0015: Cross-rates divide through a pivot currency”

Status: Accepted · Date: 2026-10-19

The obvious way to price A → B is to ask the source for the pair and use the quote it returns. That is what the service did at first: fetch with base = A and read rates[B].

Sources round their quotes. Frankfurter keeps 5 decimal places, whatever the magnitude, so a small quote arrives already starved of significant digits:

Pair Direct quote Computed through the pivot Error on 1M
JPY → USD 0.00632 0.00631951 0.49 USD (0.008%)
KRW → USD 0.00075 0.00075217 2.17 USD (0.288%)
IDR → USD 0.000055 0.00005542 0.42 USD (0.763%)

KRW → USD is not an exotic corner: it is a top-ten traded currency, and the error is twenty times the rounding the kit bothers to control elsewhere. The direct quote also makes the two directions of a pair disagree — rate(A→B) × rate(B→A) drifts away from 1 by the rounding of both quotes.

Store every rate against one pivot currency (USD by default, configurable through useCurrencyConverter({ pivot }) or CurrencyConverterService.create({ pivot })) and resolve each pair by division:

rate(A → B) = pivot[B] / pivot[A] // the pivot itself counts as 1

The cache holds one snapshot instead of one per base, and useGetRate / useConvert share it. useGetSnapshot() and useRefreshRates() therefore take no base: there is exactly one snapshot to read or refresh.

Better. The rounding sits on the well-conditioned side of the pair, so small quotes are never consumed — 1 / 1329.49 carries the precision 0.00075 had already dropped. Both directions of a pair come from the same snapshot, so rate(A→B) × rate(B→A) is exactly 1. One request serves every pair in the app, which makes the cache and the single-flight path simpler.

Worse. A pair the pivot quotes well loses a little to the second rounding: EUR → GBP becomes 0.842053 where the source’s own EUR-based quote is 0.84205 — 0.0004%, far below the ECB-vs-market drift any consumer already accepts. Choosing pivot: "EUR" for a euro-centric app removes even that.

Accepted trade-off. A provider that quotes one pair accurately and the rest poorly cannot be asked for that pair directly; the port fetches a snapshot against the pivot. That is deliberate — the accuracy of a pair must not depend on which one the provider happens to quote well.