How we grade 174,989 earnings calls, and why we don't publish buy/sell
Everything on our stock pages comes from one experiment: an AI read every earnings-call transcript in our corpus, 174,989 calls across 9,321 companies, and answered the same 500 questions about each one. This page explains exactly what we measured, what predicted the future, and what didn't.
The 500 questions
For every call, the model runs the same 500-question battery. It covers the core business verdicts (guidance, demand, margins, backlog, pricing, headcount and capex), behavioral readings such as candor, specificity, evasion, confidence, uncertainty and stress, plus hundreds of detailed yes/no checks. Applying the identical blood test to every call is what makes companies comparable across a decade.
The call grade (A to F)
The grade is a transparent composite of call quality: candor + specificity + (9 − evasion) + (9 − stress), plus bonuses when the call resolves more doubts than it creates, confidence is proportionate to evidence, a skeptic would be reassured, and guidance was raised, with penalties for critic ammunition and lowered or withdrawn guidance. Letters are fixed cuts over the full corpus: A is the top 10.5% of all calls graded, F the bottom 9.2%. A grade describes the quality of the conversation, not the stock.
The expected move (±%)
The one predictive signal that survived out-of-sample testing is about magnitude, not direction: a stock's immediate reaction to its call, combined with its pre-call volatility, predicts how far it is likely to travel over the next 45 trading days, but not which way. We publish that estimate as "expected move ±X%". It is a volatility forecast, useful for position sizing and options context, and it is the strongest claim our data honestly supports.
What failed, and why we tell you
We tested more than 1,600 directional hypotheses: split-half validation, shuffled-label placebo runs, temporal gates across eras, pre-registration. Our best candidate, a "clean call plus growing customer prepayments" pattern we call C1, beat the market in six separate backtest years and passed every statistical control. Then we froze a real forward watchlist and waited. It returned −12.4% median versus SPY. The backtests were also contaminated in a subtler way: large language models can partially remember what happened to famous stocks, so historical "predictions" about well-known names score far better than they deserve. That discovery is why no honest AI backtest on public companies should be taken at face value, ours included.
This is why our pages carry grades and expected moves instead of buy/sell calls on individual stocks. The one place we do say "buy" is the Monthly Four, a rule fixed before the buy price is known and scored in public with the median and hit rate shown. The complete story of the failure, with every table and the memorization evidence, is in Your AI Backtest Is Lying to You.