01What happened
The sovereign curve Celsmar shows covers 54 countries, from one month to fifty years. It is read from a table published by a market data provider and rendered on the bond desk.
On 4 September 2026 that table put the Turkish 10-year at 34.365%, a move of 261.5 basis points on the day. The same bond, read at another house at the same hour, was at 31.79% and had moved 4 basis points. Both agreed on the previous close: 31.75%.
One of the two was wrong that day. And here is the uncomfortable part: with a single source there was no way to know which. The figure was plausible for Turkey, the date was that day's, and the row's internal arithmetic agreed with itself. There was nothing broken that a program could detect by looking at that number alone.
02Why auditing the reader would not have helped
When data arrives from someone else's table, the natural thing is to audit the reader: check the columns pair with the right maturities, that the decimal separator is interpreted correctly, that a dash does not become a zero, that the country in the row is the one the header claims.
We do all of that, and all of it would have shown green that day. Because a reader audit answers one specific question: are we correctly reading what the source writes? And the answer was yes. It was reading a wrong number perfectly.
The missing question is a different one: is what the source writes true? That one cannot be answered from inside. It takes a second reading of the same bond, taken by somebody who is not the first.
03Where to put the threshold, and why it is not a round number
Two houses reading the same bond never agree to the decimal: they close at slightly different times, use different calculation conventions and sometimes quote different benchmark issues for the same maturity. So «they disagree» cannot be the criterion: you have to know how much disagreement is normal.
We measured it over the 83 bonds that were comparable across both houses that day:
- Median disagreement: 0.4 basis points.
- 90th percentile: 1.8 basis points.
- Largest legitimate disagreement: 11.5 bp (UK 2-year; the Japanese ultra-longs sit around 10).
- The two Turkish rows: 257 and 349 basis points.
The gap between the worst legitimate case and the wrong ones is more than twenty times over. A threshold calibrated on that distribution catches the error without firing on the ordinary noise between two honest providers, which is the whole point, because an alert that fires every day teaches people to ignore it.
04Why the fix is not to average the two
The tempting solution is to blend: take both readings and publish the mean. It is the wrong answer for a simple reason: the average of a right number and a wrong one is a wrong number, and worse, it is one that no longer matches either source, so nobody can check it against anything.
The published figure is not substituted either. What happens is that the row is flagged, the discrepancy is recorded with both readings and the date, and the screen says so. A contrast system does not decide who is right: it says that somebody is wrong, which is exactly the information a reader needs before acting on a 261 basis point move.
Frequently asked questions
Why does bond yield data need a second source?
Because auditing your own reader only proves you are reading the source correctly, not that the source is right. On 4 September 2026 one provider published the Turkish 10-year at 34.365% when it was 31.79%; the figure was plausible, the date was correct and the row's internal arithmetic agreed with itself, so nothing internal could have caught it. Only a second reading of the same bond, taken elsewhere, exposed it.
How much do two bond data providers normally disagree?
Very little. Measured across 83 bonds comparable at both houses on the same day, the median disagreement was 0.4 basis points and the 90th percentile 1.8. The largest legitimate gap was 11.5 basis points on the UK 2-year, with Japanese ultra-longs around 10. The two erroneous Turkish rows disagreed by 257 and 349 basis points, more than twenty times the worst honest case.
Should you average two data sources when they disagree?
No. The average of a right number and a wrong one is a wrong number, and it has the additional defect of matching neither source, so nobody can verify it against anything. The better approach is to flag the row, record both readings with the date, and say on screen that the figure is disputed. A contrast system does not decide who is right; it says that somebody is wrong.
Where do sovereign bond yields come from?
Market data providers publish tables of yields by country and maturity, read from the underlying markets. Different houses close at slightly different times, use different calculation conventions and sometimes quote different benchmark issues for the same maturity, which is why two honest providers still disagree by a fraction of a basis point on most bonds.
Measured on 4 September 2026 over the 83 bonds comparable across both providers that day. The published figure is never substituted or averaged: the discrepancy is recorded with both readings and shown on screen.
This is information and analysis, not financial advice, and it contains no recommendation to buy or sell. See the disclaimer.
