Score predictors vary widely in accuracy, and treating them as interchangeable is a mistake. For the USMLE Step 2 CK, a timed UWorld self-assessment correlates with the real score at around r = 0.85 to 0.88, while a raw question-bank percentage manages only around r = 0.6 to 0.7. Each tool also carries a systematic bias, over or under, that you have to correct for. Here is a tool-by-tool comparison of how accurate the main predictors really are and how to use them.
Key takeaways
- Predictors differ sharply: timed self-assessments are strong, raw percentages are weak.
- UWSA 2 has the highest single correlation, around r = 0.85 to 0.88, with a slight optimistic bias.
- NBME forms are reliable anchors but tend to run below the real score.
- Third-party predictors quote high accuracy, but are unofficial and should be sanity-checked.
- A predicted number matters less than knowing which concepts to fix.
What is a score predictor, and what makes one good?
A score predictor estimates your real exam score from your practice performance. What makes one good is two things: correlation, how tightly it tracks real scores, and bias, whether it systematically over or underestimates. A high correlation with a known bias is more useful than a loose predictor with no clear pattern, because you can correct for a known bias but not for noise. The best predictors are timed assessments built to mirror the real exam and taken under real conditions late in preparation.
NBME self-assessment forms
NBME forms are the reliable anchor for the USMLE. They correlate well with real scores and tend to run a little below the real result, commonly by around 5 to 12 points depending on the form and timing, with the newer forms more accurate than older ones. Because they are built by the same organization as the exam, they are trusted as a backbone predictor, and averaging two recent forms is usually more credible than any single non-NBME number. They cost around $60 each, and are best sat under timed conditions in the final weeks.
UWorld self-assessments (UWSA 1 and 2)
UWorld's self-assessments are heavily used and heavily mythologized. UWSA 2 is often the single most powerful predictor, correlating with the real Step 2 CK at around r = 0.85 to 0.88, but it carries a slight optimistic bias of roughly 0 to 5 points, so a predicted 245 may correspond to a real 241 to 245. Take it 7 to 14 days out under full timed conditions and read it as a ceiling rather than a floor. UWSA 1 tends to overpredict more and is better used for mid-preparation confidence and stamina than as a final forecast.
Raw question-bank percentage
This is the predictor students overvalue most. Cumulative question-bank percentage correlates with the real Step 2 CK at only around r = 0.6 to 0.7, and it is contaminated by block order, timed versus untimed sessions, first pass versus review, and uneven subject spread. Two students at the same percentage can finish 15 points apart. Use your percentage as a learning signal, not a score forecast, and do not let a comforting number override a timed assessment.
AMBOSS and third-party predictors
AMBOSS includes its own Score Predictor within its platform, and several third-party tools convert NBME, UWSA, and question-bank inputs into a predicted score. Some third-party predictors quote very high accuracy, such as correlations around r = 0.9 with most predictions within a handful of points, based on large collected datasets. Treat these claims with reasonable caution: they are unofficial, their datasets and methods vary, and they are best used to sanity-check the official NBME and UWSA numbers rather than to replace them.
How to use predictors sensibly
Combine them and correct for bias. Use recent NBME forms as your backbone, weight UWSA 2 heavily while trimming a few points for its optimism, and treat your question-bank percentage as background rather than a forecast. Take predictive assessments late, under exam-like conditions, and accept the realistic band: even good predictors land within about 5 to 8 points, not on the nose. Once two or three recent scores agree, more numbers add anxiety, not information.
The deeper point: a number is not a plan
Even a perfect prediction does not tell you what to study next. The actionable question is which concepts you keep missing, and that is where a predicted score falls short. iatroX approaches this from the other end, using semantic adaptive learning to map the underlying weaknesses behind your errors so your study targets the gap rather than chasing a number, across UK and international exams, at £29 per month or £99 per year with free sample questions. A forecast tells you where you stand; mapping your gaps tells you what to do about it. Try the free questions, and read the companion piece on practice scores versus the real exam.
Frequently asked questions
What is the most accurate USMLE score predictor? For Step 2 CK, UWorld Self-Assessment 2 has the highest single correlation, around r = 0.85 to 0.88, with a slight optimistic bias. Recent NBME forms are also strong and tend to run below the real score.
Are NBME forms accurate? Yes, they are the reliable anchor, though they tend to underpredict by around 5 to 12 points depending on the form and timing. Newer forms are more accurate, and averaging two is more credible than one.
Is my question-bank percentage a good predictor? No. It correlates only around r = 0.6 to 0.7 and is biased by how you used the bank. Use it as a learning signal, not a forecast.
Can I trust third-party score predictors? Use them cautiously. They quote high accuracy but are unofficial with varying methods. Sanity-check them against official NBME and UWSA numbers rather than relying on them alone.
How accurate can any predictor be? Even good predictors land within about 5 to 8 points of the real score, not exactly. Use several, correct for known bias, and take them late under real conditions.