Two numbers, kept apart on purpose. Merging them would be the easiest way to mislead you here.
Trained on past fiscal years, then tested against a year it had never seen. Repeated across folds. This is real, but it is a simulation of the product rather than the product itself.
| What we had to go on | Correct on confident calls |
|---|---|
| Same buyer, same category, same industry code | 92.7% |
| Same buyer and category, any industry code | 80.0% |
| Same category and industry code, any buyer | 80.2% |
| Same buyer only — the thinnest evidence we will act on | 82.8% |
Every prediction this system publishes is archived before the solicitation closes. When the award later appears in federal records, we look up how many bidders it really drew and write down whether we were right.
Not enough yet to publish a rate.
0 predictions archived · 0 resolved against real awards so far. We will not quote a live percentage until at least 200 have resolved, because a rate built on a handful of results is noise wearing a decimal point.
Federal awards take weeks to months to post after a solicitation closes, so this section fills in slowly. It will be empty for a while, and then it will be the most useful thing on this site.
The archive is append-only, enforced by the database itself. Once a prediction is written, what we said cannot be edited and the row cannot be deleted — only the outcome can be added later. There is no path, including for us, to quietly improve a past call after seeing how it turned out.
One limit worth stating: federal records carry a solicitation number on roughly 58% of definitive contracts, so a share of our predictions will never be matchable to an outcome. Those stay unresolved rather than being counted either way.