Your Screener Is Lying to You

In 2023, researchers from Penn State and Columbia published a study in the Journal of Accounting and Economics titled “Lost in Standardization.” They compared the raw financial data in SEC filings against the same data as reported by Compustat, the S&P Global database that powers most of the financial tools retail investors use.

Their findings were stark: 68 of 73 commonly-used financial data items showed statistically significant discrepancies. Revenue alone differed in roughly 26% of the companies they sampled. And Compustat wasn't even the worst offender — FactSet data showed even larger deviations from the original filings.

A separate study titled “Using XBRL to Conduct a Large-Scale Study of Discrepancies” published in the Journal of Information Systems found similar massive gaps, demonstrating that data providers frequently alter reported numbers to fit uniform templates.

In NVIDIA's fiscal 2025 10-K, the company reports $130.5 billion in revenue. A typical screener shows you this as a single “Revenue” line, maybe split into two segments. But the actual filing tells a much richer story.

According to Palantir's 2024 10-K filing, the company reported $2.87 billion in revenue with $692 million in stock-based compensation — 24% of revenue. This is one of the most important numbers for valuing Palantir, because SBC represents real dilution to shareholders.

In Palantir's actual 10-K, stock-based compensation is broken out as a dedicated line and further detailed in the notes, showing how it's distributed across cost of revenue, sales and marketing, R&D, and G&A. This distribution matters: SBC concentrated in R&D signals investment in future product; SBC concentrated in sales signals customer acquisition cost.

A typical screener shows you a single “Stock-Based Compensation” number in the cash flow statement — if it shows it at all. Some roll it back into the operating expense categories, making it impossible to separate real cash costs from non-cash dilution. The functional breakdown that tells you where the company is spending equity disappears entirely.

The “Lost in Standardization” researchers tested something specific: do these data discrepancies actually change investment conclusions? They examined the accruals anomaly, one of the most well-documented patterns in finance — companies with high accruals tend to have lower future stock returns.

When they calculated accruals from the original SEC filings, the anomaly was statistically significant. It worked as a return predictor. When they calculated the same metric from Compustat data, the anomaly disappeared entirely. The standardization process had introduced enough noise to destroy a real, exploitable signal.

This isn't a theoretical concern. If you're building a DCF model, screening for quality metrics, or comparing margins across an industry — and you're using data from a typical screener — the numbers you see may not support the conclusions you draw from them.

DeepFundamental parses SEC filings directly and preserves the company's own reporting structure. We don't force NVIDIA's four revenue segments into one line. We don't collapse Palantir's SBC detail into a generic bucket. Every line item traces back to the original filing.

The goal isn't to replace screeners for quick filtering. The goal is to give you the actual data when you sit down to do real analysis — so the numbers in your model match the numbers the company reported under oath.