How the model has actually performed

Every stock on this site carries a score, and the strongest of them get a Top 1%, Top 5% or Top 10% badge. This page is the scoreboard for those badges. For each month it asks one question: if you had bought the badged names at the next morning's open and held them for 30 trading days, how did that compare with simply owning the whole universe?

Read the status column first. Months marked unseen are the real test: the model had never seen any part of them, not in training and not in the checks used to pick which version to ship. Months marked holdout were not trained on either, but they were used to choose the final model, so treat them as the more flattering half of the record.

Long, month by month

Month	Status	Top 1%	Top 5%	Top 10%		Whole market
Loading…

Top line: the median 30-trading-day return of the names carrying that badge during the month, entered at the next open. Second line: how many names carried that badge, and how many of them actually ended higher, which is what happened rather than what the model predicted. "Whole market" is every scored name that month, which is what the badges have to beat to be worth anything.

Short, month by month

The same model also marks the weakest names Bot 1%, Bot 5% and Bot 10%. This table asks whether shorting those would have paid. It is written as the short's own profit, so a positive number means the short made money, and it is compared against shorting the whole market rather than against owning it.

Read this one as a negative result. The bottom bands are not a short signal. They made money only in the months the entire market fell, and barely more than shorting the index would have, while winning under half the time. The deepest band has been the worst of the three. Published because it is what the measurement says, not because it is a strategy.

Month	Status	Bot 1%	Bot 5%	Bot 10%		Short the market
Loading…

Top line: the median 30-trading-day profit of shorting at the next open. Second line: how many names carried that badge, and how many of them actually ended lower, which is what a short needs and, again, is what happened rather than what the model predicted.

What is and is not included

Only complete months. A month appears once every one of its bars has a full 30-day forward window in the data, so a partly-finished month can never flatter or spoil the record. The current month is therefore always absent, and the newest month arrives about six weeks late.
Medians, not averages. One takeover bid can double an average and say nothing about the other ninety-nine picks. The median is what a typical pick did.
Taken blind, so half of them did better. Every name carrying a badge is counted here, bought at the next open and held exactly thirty trading days, with no judgement applied at any point. Nothing was skipped for looking wrong, no entry was timed, no loser was cut and no winner was left to run. A median is the middle of that: half the names in a band finished above the figure shown. None of these numbers measure choosing between the names in a band or reading the chart before buying, so they are a floor for someone doing neither, not a ceiling on what the badges can find.
High-volatility names are excluded, exactly as they are on the site. They never receive a badge, because the model was never trained on stocks that volatile.
No trading costs. These are raw price returns, so a real result after spread and commission would be lower.

This is one model's ranking measured against the market it ranked, not a claim about what any individual stock will do next. A month where the badges beat the field is evidence the ordering carries information. It is not a promise about the month you are in.