USMLE Score Prediction Methodology: The Science Behind the Forecast

Predictive Modeling in Medical Education

A score predictor looks at the practice scores you enter and asks what Step score tends to go with them. Ours do that for Step 1 (as a pass probability), Step 2 CK and Step 3 (as a 3-digit estimate with a range). The answer can only be as good as the data and assumptions behind it, so we describe both plainly below.

What Each Tool Is Built On

The quality of an estimate depends on its data, so here is what sits behind each tool. All three are calibrated against real score reports shared by students (next section).

• Step 2 CK: a reference dataset of 5,039 records (practice scores alongside Step 2 CK scores). Inputs: NBME Forms 9–16, UWSA 1–3, Free 120 % and UWorld %.
• Step 3: a reference dataset of about 4,500 records, including 1,216 student-reported results. Inputs: Step 2 CK score, UWorld Step 3 %, Free 137 %, UWSA 1–2 and NBME CCMSA 6, 7 and 8 (% correct only).
• Step 1: no outcome dataset inside the tool. It uses our own approximation of how NBME CBSSA scores (Forms 29–34) relate to the probability of passing, with Free 120 and UWorld % as supporting inputs. NBME does not publish such a table — the official probability of passing is printed on each examinee’s own NBME score report, and that is the number to trust.

How We Calibrate

We calibrated the calculators against 4,758 score reports that students shared publicly, most of them between 2024 and 2026: 1,742 for Step 1 (including 294 fails), 1,001 for Step 2 CK and 2,015 for Step 3. We fitted the Step 2 CK adjustments on older reports (before July 2026) and checked every change on newer ones (July to October 2026). The Step 1 curve is our own approximation and was not tested this way, and the Step 1 NBME 34 adjustment is our own, with too few reports to test it. The numbers below come from that newer set.

What the real reports changed:

• Step 2 CK: NBME Forms 9–16 scores come out about 4 points higher in the model, UWSA 1–3 scores are mapped onto the Step 2 CK scale, and the 80% range is set by how many practice scores you enter (wider for one score, narrower for many) to match how far real scores land from the estimate.
• Step 3: NBME CCMSA scores (Forms 6, 7 and 8, entered as % correct only) now move the estimate.
• Step 1: NBME 34 is accepted with a downward adjustment of up to 5 points (our own adjustment; there are too few Form 34 reports to test it). The curve itself is our own approximation, because NBME publishes no conversion table.

How it performed on the July–October 2026 reports, as the average gap between the estimate and the real score:

• Step 2 CK: typical error about 8 points, from about 9 before calibration.
• UWSA score converter: typical error down from about 13 to about 7.5 points.
• Step 3: typical error about 9 points.

Two things to keep in mind. These are self-reported scores from students who chose to post them, so they are not a random sample. The tool flags only a few of the students who end up below the passing standard (3 of 17 in our newer reports), so if your estimate or practice scores are below about 240, take a fresh NBME close to your exam and compare it with the passing standard of 218.

The Statistical Framework

We move beyond simple averages. Each Step 2 CK prediction is a blend of three methods — a 3-method ensemble (K-Nearest-Neighbor matching, weighted averaging, and per-form regression) — so no single signal dominates:

• Nearest-neighbor matching (40%): finds the records whose practice scores look most like yours and averages their Step 2 CK scores, giving closer matches more weight. Records beyond a fixed similarity cut-off are ignored.
• Weighted averaging (35%): combines your scores using fixed per-form weights — newer NBME forms, UWSA 2 and the Free 120 count most — and adds the dataset’s average gap between practice scores and recorded Step 2 CK scores.
• Per-form regression (25%): adjusts each form you entered by that form’s average gap between practice score and recorded Step 2 CK score, so a form that reads high or low in the data is corrected for.

Accuracy: What We Can and Can’t Claim

The typical errors above describe past reports, not a promise for any one student. Each Step 2 CK result also shows a range. The range is fitted so that about 80% of held-out records fell inside it. Real results can land outside it, so read the whole range, not just the middle. The range is wider when you give the tool less to work with.

For a sense of scale, USMLE’s own figures for the real exam: the standard error of measurement is about 6 points for Step 2 CK and 5 points for Step 3, and two scores need to differ by about two standard errors of difference (roughly 16 points on Step 2 CK) before they likely reflect different proficiency (USMLE Score Interpretation Guidelines, updated 17 August 2026). No prediction can be more exact than the exam it predicts. More detail on each tool: Step 2 CK, Step 1 and Step 3.

Data Privacy & Ethical Standards

We are committed to the ethical use of analytics in education.

• No account: the tools never ask for your name or email. The practice scores you enter are used to generate your prediction.
• Optional score sharing: if you share your real Step 2 CK score after the exam, it is recorded anonymously by our analytics provider (Umami). It is not added to the model automatically. Our usage analytics (Umami, PostHog) never receive the practice scores you type or the score we predict.
• No selling: we never sell student data or share it with residency programs or institutions.
• Transparency: you should understand the why behind your estimate, which is why this methodology — including its limits — is published openly.