Loading...
Please wait while we set up your experience
Every assumption, coefficient and limitation.
We start from the published offer rate for your course. Then we shift it in log-odds space using factors that admissions data shows are predictive: grades, school type, TMUA score, STEP plan (Cambridge Maths), personal statement strength, interview prep, olympiad results, contextual flag, and college choice for Oxbridge. Each shift is small. None of them is fate. The largest single factor is the admissions test, which can move probability by up to 20pp. The smallest is interview prep at around 3pp. For Oxbridge we use a two-stage model: P(offer) = P(shortlist) x P(offer | shortlist). For Cambridge applicants we also surface the pool: ~700 applicants are pooled each year, ~22% of them get an offer from a different college. Around every estimate we display an 80% confidence band that reflects cycle-to-cycle variation, factors the public data does not capture, and small samples at niche courses. Source citations sit under each factor in the result card.
Uni Predictor is a logistic / Bayesian-style econometric model. All factors compose multiplicatively in log-odds space rather than as multiplicative percentages, which:
For Oxbridge we use a two-stage model: P(offer) = P(shortlist | application) × P(offer | shortlist). For non-interviewing universities (Imperial Maths and Computing, LSE, UCL) we use a one-stage model: P(offer | application) where the admissions test plus predicted grades carry the most weight.
~700 Cambridge applicants are pooled each year. ~22% receive an offer from a different college. Our model computes P(combined) = P(direct) + P(rejected directly but pooled) × P(pool offer). Pool entry probability is calibrated to peak around the direct-offer rate; very strong and very weak applicants pool less. Cambridge Maths pools at ~25%, Computer Science at ~22%, Engineering at ~21%, Economics at ~18%. These are 2024-cycle averages.
A logistic, Bayesian-style econometric model. All factors compose multiplicatively in log-odds space rather than as multiplicative percentages, which avoids the “stack to 100%” trap, lets us compose adjustments cleanly without hard ceilings, and propagates uncertainty so variance on each adjustment sums in log-odds.
For Oxbridge we decompose P(offer) into P(shortlist | application) × P(offer | shortlist). For non-interviewing universities (Imperial Maths and Computing, LSE, UCL) we use a one-stage model where admissions test plus predicted grades carry most of the weight.
Course-specific log-odds curves derived from typical successful applicant ranges. TMUA weights are calibrated slightly higher than published research alone supports, reflecting direct conversations with current Cambridge maths tutors and recent students. We recalibrate when Cambridge Assessment publishes new TMUA outcome data.
Around 700 Cambridge applicants are pooled each cycle. Around 22% of them receive an offer from a different college. The model computes P(combined) = P(direct) + P(rejected directly but pooled) × P(pool offer). Pool entry probability peaks around the direct-offer rate. Cambridge Maths pools at around 25%, CS at 22%, Engineering at 21%, Economics at 18% (2024 averages).
An 80% confidence interval that captures cycle-to-cycle variation, factors the public data does not measure (interview performance, personal statement quality), and small-sample noise at niche courses. Predicted grades widen the band by 1.4× compared to achieved grades, reflecting the 16% UCAS prediction-accuracy figure from Wyness (2016).
We use the Boliver (2013) odds ratio of 0.59 for state vs independent at equivalent grades, blended with each university's own grade-controlled gap (per OfS 2024). Cambridge grammar-school applicants actually accepted at 24.4% in 2024, ahead of independent at 21.6%, so we bake the real direction in rather than the headline assumption.
Oxford only publishes college-level all-subject rates, not per-subject. Our Oxford college multipliers blend the all-subject signal with the small number of college-level subject feedback PDFs the Mathematical Institute and CS Department do publish. Confidence bands widen for Oxford colleges where per-subject data is thin.
TMUA-specific predictive validity has not been peer-reviewed. Predicted-grade inflation is systematic (only 16% accurate per Wyness 2016). Personal statement weight is qualitative. Oxford per-college, per-course data is partial. Oxford 2026 TMUA cycle is the first ever, so thresholds are extrapolated from MAT-era data and Cambridge cohorts.
v2.0 (Uni Predictor) calibrated 2026-05-19. The model recalibrates whenever new End-of-Cycle data or FOI returns arrive.
Ready to see your number? Open the Uni Predictor and drag your TMUA.