How Vero works
What goes in, what happens, and how well it works.
Vero is a research workspace for long-term investors and active traders. This page names every source, explains each method in plain words, and gives the measured accuracy against simple baselines, including where it falls short.
Inputs
Four licensed or public sources, named.
IEX Historical Data
End-of-day prices and volume built from trades on the IEX exchange, back to December 2016.
Next morning (T+1)
SEC EDGAR
Company filings (8-K, 10-Q, 10-K), reported financials, company names and industry codes.
As filed
GDELT Project
News discovery: the headline, the publisher and the link. Vero does not have the article text.
Through the day
Official calendars
Release dates published by the Federal Reserve, BLS and BEA: FOMC, CPI, jobs, GDP.
As published
Process
Three engines, each checked against something simpler.
The thesis engine
You write down why you own a stock. Vero splits it into claims that could turn out true or false. Each new filing or scored headline on that company is read against those claims, and the engine returns one of three verdicts (confirms, challenges or breaks) naming the claim and the reason, or no verdict when the item bears on none of them.
It judges only from the text in front of it: a headline alone rarely carries a verdict, and the company's own filing is read from its text.
The forecast (1 week and 1 month)
A range, not a price target: the 50% and 80% bands a stock's close is likely to fall in, from a volatility model that knows about upcoming earnings dates. It has no direction signal, and the middle line is flat.
It is tested walk-forward: every trading day since late 2019 is a forecast made only with data known at the time, refit every 21 sessions. A horizon is shown only if it beats two random-walk baselines, in at least 90% of bootstrap resamples, with bands that hold the share of outcomes they claim.
The events service
Upcoming earnings dates are estimated from each company's own SEC filing history, and shown as "around" a date because they are estimates until the company announces them. Economic dates come from the agencies' published calendars, with a link to each.
One real output
A filing, a thesis, a verdict.
Real engine output · NVIDIA (NVDA)
1 · The news item
NVIDIA files 8-K: Results of operations (earnings release) (Item 2.02)
SEC EDGAR · filed Aug 26, 2026
NVIDIA reported its Q2 FY2027 results, with revenue of $96.2 billion, up 106% from a year ago, and data-center revenue of $89.0 billion. Gross margins were 75%, GAAP EPS was $2.46 and non-GAAP EPS was $2.22. The company also outlined a strong Q3 outlook of about $108 billion in revenue and highlighted Vera Rubin platform ramp and other AI infrastructure momentum.
2 · The thesis it is checked against
Example thesis written for this demo. Tap a claim to see what it assumes.
3 · The verdict, and why
confirmsGross margin stays above 70%NVIDIA's reported Q2 FY2027 gross margin was 75%, directly supporting claim c2 that gross margin stays above 70%.
The engine's own words, unedited.
Example output generated on 2026-09-30. Real output of Vero's thesis engine on NVIDIA's 8-K filed Aug 26, 2026, judged against an example thesis written for this demo. Not advice.
Accuracy
Measured against baselines, not against hopes.
| Horizon | Error (CRPS): Vero / naive / strict | Improvement vs naive / strict | 80% / 50% bands held | Status |
|---|---|---|---|---|
| 1 week | 2.445 / 2.511 / 2.503 | +2.63% / +2.31% | 79.3% / 48.7% | Shown |
| 1 month | 5.129 / 5.219 / 5.169 | +1.71% / +0.76% | 77.8% / 47.9% | Shown |
| 3 months | 8.592 / 8.762 / 8.544 | +1.94% / −0.56% | 77.1% / 46.9% | Not shown: fails the strict baseline |
- CRPS scores the whole forecast range against what happened; lower is better. "Naive" is a random walk using the last year's volatility. "Strict" is a random walk using the better of 3-month and 1-year volatility, with the same fitted shape as Vero's.
- The edges are small, and they are stated as they are: a few percent better than a random walk at 1 week and 1 month. At 3 months the model does not beat the strict baseline, so no 3-month forecast is shown.
- Direction: the middle line has no view on up or down. Its hit rate matches "always up" (1 week: 53.8% vs 53.7%).
- Per company, a forecast is shown only where that company's own record passes: 49 of 49 at 1 week, 40 of 49 at 1 month.
Earnings-date estimates
Walk-forward on 3,278 past releases from 47 companies' SEC filings: median error 1 day, 62.5% within a day, 92.2% within 7 days (mean error 3.5 days). That is why Vero shows an estimated date as "around" it.
Thesis verdicts
There is no accuracy number for verdicts yet: whether a claim was "really" confirmed is a judgment, and Vero has not published a scored record. Every verdict shows its reason and links its source so you can check it yourself.
Limits
What Vero does not do.
End of day, the next morning
Prices are yesterday's close (T+1). There are no live or intraday quotes.
IEX-only volume
Volume counts trades on the IEX exchange only, a small share of all US trading.
Estimated earnings dates
Until a company announces its date, Vero's date is an estimate from its filing history.
AI can be wrong
Summaries, verdicts and assistant answers come from language models. Each links its source; check it.
Headlines, not articles
For news found via GDELT, Vero has the headline and link, not the article text.
Not advice
Vero explains and organizes information. It never tells you to buy, sell or hold, and places no trades.
Start free
See it on your own companies.
The free plan covers 3 stocks, with no card and no end date.