Next-Unit English Writing Scores under Repeated Cohort Transfer

Abstract

A single future cohort can make an 80% prediction interval look stable. This paper predicts the next English writing score from information available at the current unit, then repeats the same train–calibrate–test transfer on seven successive entering cohorts. Scores are converted to a percent of that year’s full mark before modeling. The de-identified records are English writing assessments from ten undergraduate cohorts during 2016–2025 at the School of Mathematical Sciences, Inner Mongolia University. Across the seven transfers, last-score MAE is 5.40–5.68 points. Linear quantile P50 lowers it to 4.81–5.15, and quantile LightGBM to 4.80–5.30. Raw LightGBM P10–P90 intervals cover 69.4%–76.7% of outcomes at mean widths of 12.9–14.6. One pooled conformal correction moves coverage to 77.8%–82.0% and width to 15.1–16.6 on the six transfers whose calibration cohort has more than 200 students; the remaining transfer, calibrated on 121 students, reaches 83.8% at width 17.2. A separate correction for each target unit changes LightGBM coverage by at most 1.7 percentage points. The raw linear interval covers 77.5%–81.1%. Score history accounts for most of the point-error reduction. A student-clustered bootstrap does not find a year in which LightGBM has a lower MAE than the linear quantile. A short explanation interface repeats these numbers under a fixed contract. We did not collect teacher evaluations of the interface.