The same ratio shows different values across data providers because each provider makes its own choices about which price, which earnings figure, which period and which definition to use, and most of them do not tell you which choices they made. Usually none of the numbers is wrong. They answer slightly different questions under the same label. For banks and asset managers, the fix is not to find the "correct" provider. It is to write down which variant the firm means and insist that every source, human or machine, shows how it got there.
Price → Earnings → Period → Definition
That sequence is how I take apart any ratio that does not match what I expected. In my experience, almost every discrepancy traces back to one of those four steps.
A problem every analyst learns to live with
Over my career in finance, I watched the same ratio take different values depending on where I looked. On a Bloomberg screen, a P/E might be built on trailing earnings in one place and forward earnings in another. One field might use GAAP earnings and the next adjusted earnings. Experienced analysts learn to check how a number was built before they rely on it. Junior analysts often learn this the hard way, when a figure in a pitch book does not tie to the figure in the credit memo.
What I had treated as a known nuisance became much clearer when I saw a piece of work by Jeroen Bouma, the builder behind the open-source Finance Toolkit library, whose career has run through pension fund management and product leadership at OpenBB. He took one company, Microsoft, and one ratio, P/E, and found eight different values. These were not eight errors. They were eight defensible calculations, each produced by a different combination of reasonable choices.
When we discussed it on Stack to Street, Jeroen made the point that this is not peculiar to P/E. Whether you look up price-to-earnings, return on equity or price-to-book, "all the different websites report a different number." He was careful to add that most of them can still be correct. I think that is the part executives should take seriously. The problem is not bad data. The problem is undisclosed definitions.
Walking the four variables
Price
The numerator looks simple until you ask which price. Jeroen put the question directly: are you going to use the price of yesterday, today or last week? For a large, liquid stock in a calm market, the difference may be small. For a smaller name, or in a volatile week, it can move a valuation multiple enough to change a screening result. A provider that refreshes intraday and one that uses the prior close will disagree all day, and both will be accurate.
Earnings
The denominator is where the real divergence lives. Forward estimates and reported results tell different stories. Consensus forward earnings reflect analyst expectations. Trailing earnings reflect what actually happened. On top of that sits the GAAP versus adjusted split, where one-off charges, stock-based compensation and restructuring costs may or may not be stripped out. Two providers can agree on the price to the penny and still be far apart on P/E because one is dividing by a cleaner, management-adjusted figure.
Period
Even once you have chosen trailing GAAP earnings, you still have to choose the window. Jeroen mentioned using an average of the last couple of quarters as one option. Others sum the last four reported quarters, use the last fiscal year, or annualize the most recent quarter. Each approach treats seasonality and recent momentum differently. For a company whose earnings are inflecting, the period choice alone can produce a materially different multiple.
Definition
The last step is the formula itself: diluted or basic share counts, how negative earnings are handled, whether minority interests are included. This is where providers most often go quiet. You see a number and a label, and you are left to infer the recipe.
There is a further variable that sits underneath all four, and Jeroen learned it from his users. Two developers in Italy building a trading model on the Toolkit pointed out that financial statements arrive in different currencies and cannot be compared directly until they are converted to a common one. It sounds obvious once said. In practice, a global equity screen that mixes reporting currencies without a consistent conversion rule will rank companies on figures that were never comparable. For institutions with cross-border portfolios, I would add currency treatment to the definition step explicitly.
Why writing it down is the actual answer
Jeroen's response to the eight P/E ratios was not to declare one of them correct. It was to write down every variation. He started Finance Toolkit in 2019, well before the current wave of AI tooling, as a personal learning project. He looked up each ratio, worked out its variations, and recorded them in code. It began with roughly a hundred financial ratios and grew past 500 metrics, extending into DuPont decomposition, GARCH models, Fama-French factors, value at risk and technical indicators.
I find that origin instructive for institutions. The discipline that produced the library was to refuse to use a metric without understanding and documenting how it was built. Most firms ask their analysts to do exactly that. Few make the analysts write it down in a form anyone else can inspect.
His reason for publishing the formulas openly also speaks to the core problem. If he wrote all of them down and kept them private, he said, he would be no different from every other website that calculates a metric without exposing how. "Then I'm just one of the many." I agree with the logic, whether or not a firm uses open-source code at all. A number whose method you cannot see is just one more number that disagrees with the others. A number whose method you can read is something you can reconcile.
What this means for banks and asset managers
Most institutions already pay for one or more terminals and data feeds. The discrepancy problem does not go away by consolidating onto a single vendor. It simply becomes invisible until someone compares that vendor's output against a counterparty's model, a rating agency's figure or a company's own investor presentation. At that point, the firm needs to know which variant it used and why.
I see three practical consequences.
First, definitions are a governance matter, not a preference. When a credit committee approves a deal on a leverage multiple, or a portfolio manager screens out names above a P/E threshold, the definition behind that number has become part of the decision. It should be recorded like any other input.
Second, reconciliation should start with the recipe, not the result. When two numbers disagree, teams often spend hours checking whether one feed is stale. Walking the four variables first usually resolves the gap in minutes.
Third, the problem is accelerating. Analysts now ask AI tools to produce ratios, and a general-purpose model makes its own choices across price, earnings, period and definition without saying so. I have written separately about how to stop AI tools from guessing how a metric is calculated, including routing requests to documented calculations. The underlying point is the same one this article makes: if the definition is not written down somewhere a machine can follow, the machine will improvise. In markets with even less standard vocabulary, the stakes are higher still, which is why I think private credit could benefit from shared metric definitions.
A checklist for finance executives
- Name the house variant for every decision-critical ratio. Specify price date, earnings basis (trailing or forward, GAAP or adjusted), period window and share count. Put it in writing.
- Add currency treatment to the definition. State the conversion rule and rate source for any cross-border comparison.
- Require providers to disclose methodology. For each metric you rely on, ask the vendor for the formula and the inputs. Treat "proprietary" as a risk flag, not an answer.
- Reconcile by recipe first. When numbers disagree, walk Price → Earnings → Period → Definition before escalating to data quality.
- Keep a reference implementation. Whether built in-house or drawn from inspectable open-source code, maintain one calculation the firm can reproduce on demand.
- Extend the same standard to AI tools. Any model that produces a ratio for an analyst should be held to the same disclosure test as a terminal.
The number on the screen will keep varying from provider to provider. What institutions control is whether they can explain, every time, which version they used and why it was the right one for the decision in front of them.
About the author
Michael McLaughlin has more than 30 years in structured finance and private credit. At Macquarie Group from 2008 to 2023, he built the credit markets business from inception into a portfolio financing platform that averaged a $5 billion portfolio and generated more than $1 billion in cumulative net revenue. He has advised Hebbia on AI research software for financial services, and builds his own finance applications with AI coding tools and open-source software.
Through Path66, Michael works with financial institutions and investment firms on asset finance, liquidity, capital raising, data-driven due diligence and the practical application of emerging technology in finance. He hosts Stack to Street, where he interviews the developers building the software that will change financial institutions. Full bio.
Source
Your AI Shouldn't Guess the Formula, Stack to Street with Jeroen Bouma, Creator of Finance Toolkit. Full transcript on the episode page.