# Before You Publish a Number, Try to Break It

- Published: 2026-10-01
- Author: [Connectively](https://blog.connectively.us/author/terkel.md)
- Topics: [Thought Leadership](https://blog.connectively.us/category/thought-leadership.md)
- Source: https://blog.connectively.us/before-you-publish-a-number-try-to-break-it/
- Publisher: Connectively

*Written by Razvan Luca*

I built Talval, an [independent research site](https://talval.com/about) that publishes valuations for about 1,000 listed companies. Almost every page is numbers, and the main thing I’ve learned this year is that wrong numbers rarely look wrong. They look reasonable, sit on the page for weeks, and nobody complains.

Here are three that taught me the most, and the rule each one left behind.

The first looked perfectly plausible. For months our page on American Express showed annual revenue of about $41 billion. That’s a big, believable figure, and it came from the line most financial data labels as revenue. It was wrong. That line only counts fees from customer contracts. The company’s own total, after interest, was about $72 billion. Every ratio built on it was off too, from margins to growth.

We only found it this week, when we added SoFi and the same trap turned $3.6 billion of revenue into $619 million. Six times too small is hard to miss. A figure 43% too low, on a company everyone has heard of, isn’t.

The fix was a rule: when a company reports a net revenue figure that’s clearly bigger than its fee revenue in every year, use the net figure. Before shipping it, we ran it across all 651 companies it could reach. It changed three: American Express, Interactive Brokers and SoFi. Everything else stayed put, which is what we wanted to see.

The rule it left: test a fix on everything it could change, not just the case that prompted it. A fix that quietly moves 200 other companies is a new bug wearing a nice label.

The second was absurd, and still nobody noticed. Shares in London are priced in pence, while company accounts are in pounds. One function on our site forgot to divide by 100. For a while AstraZeneca’s annual revenue read £4.4 trillion, and Coca-Cola Europacific Partners was apparently worth £3.4 trillion. The real figures were around £44 billion and £34 billion.

No reader ever flagged it. That’s the uncomfortable part. A number can be impossible, bigger than the whole UK economy, and survive anyway, because people assume software gets arithmetic right. I did too.

The rule it left: write down what impossible looks like, then check for it automatically. A profit margin above 100% can’t happen. Revenue bigger than a country’s GDP can’t either. Those checks cost minutes to write and they catch exactly the errors a tired human skims past.

The third is a number we chose not to publish. Our valuation model sometimes lands a long way from the market price. When the gap is more than 60%, it’s more often a broken input than a hidden bargain, so we show the direction, undervalued or overvalued, and hold the figure back. Today that’s 170 of the 985 companies we analyze.

Even that cutoff needed a fix. The first version used a single threshold, and one of our most-viewed stocks was sitting at 59.99%. One cent either way in its share price would have made its fair value vanish and come back. To a reader, that looks like a model changing its mind every few minutes. Now it takes 60% to hold a new figure back and 65% to withdraw one that’s already out.

The rule it left: a hard threshold on a moving number needs a buffer, or your page will change its mind every time the market twitches.

None of these were clever bugs. Each one was a number nobody had tried to break. So before anything new goes on the site, I ask three questions. What would this look like if it were wrong? Where else does this change land? And would I rather publish nothing than this? The last one saves us more often than I’d like to admit, and it’s the one I’d suggest to anyone whose work ends in a number other people trust.

*About the Author:* *Razvan Luca is the founder of* *[Talval Research](https://talval.com)**, an independent equity research platform covering about 1,000 US and European companies.*