This study predicts word-level errors in Chinese learners’ English writing, testing whether features from first language (L1) negative transfer improve performance and interpretability. A unified dataset was built from three corpora (FCE, Lang-8, CAWSE). Models include Logistic Regression, LightGBM, BERT, and few-shot prompting. Beyond baseline linguistic features (length, frequency, POS), four L1 transfer categories were introduced: high-risk word lists, grammatical templates, tense/aspect markers, and countability cues. Results show LightGBM achieved the best in-domain scores (PR-AUC up to 0.83), while Logistic Regression benefited most from L1 features. Cross-domain tests highlight Lang-8’s weak but diverse annotations as better for generalization. Overall, L1 features yield modest but consistent gains, enhancing both prediction and interpretability for learner-centered NLP.