A startup score is a number that ranks a company against other companies. Funds and accelerators use scores to decide which applications to read first.
A score is useful only if it predicts outcomes. Thus, the most important question is not how a score is made, but how it was tested.
Two ways to make a score
Rubric scores. Partners score each deal against a written list of criteria, such as the team, the market and the traction. This is simple and clear, but different partners often give different scores.
Model scores. A model learns from many past companies and their outcomes. The same inputs give the same score every time. The model must be tested on companies that it did not learn from.
How to test a score
Use companies where you know the outcome, but hide the outcome from the model. Rank the companies by score. Then compare the success rate of the top group with the base rate of all companies.
Fix the test rules before you run the test. If you change the rules after you see the result, the result does not prove anything.
Log each live score with its date before the outcome is known. Only a live, dated record can show that a score works on new companies.
How IdeaPulley scores startups
Arlo, our own model, sets every IdeaPulley score. Arlo learned from years of real startup outcomes. An AI explains each score, but it cannot change it.
On a test of 5,983 Hacker News founders, the top 20% of our ranking held 2.62 times the base rate of companies that later raised $5M or more. No live prediction has resolved yet, and we publish that too.
Questions
What is a startup score?
A number that ranks a startup against other startups, so that a fund or accelerator can decide which applications to read first.
How do you know if a startup score works?
Test it on companies with known outcomes that the model did not learn from, compare the top group with the base rate, and log live scores with dates before outcomes are known.
Can AI predict startup success?
A model trained on real outcomes can rank startups better than chance, but no model is certain. Ask for the test method, the base rate and a dated record of live predictions.