Legal Logit Model for predicting judicial disagreement in Indian courts

Once a case reaches the Supreme Court on appeal, the justices may either affirm or reverse the judgment of the lower court. Forecasting such judicial disagreement is important not only for predicting outcomes but also for understanding the judge-specific and case-specific factors that drive these decisions. This study aimed to present the Legal Logit Model (LLM), an evolved neural network-based version of the Multinomial Logit (MNL) model. The LLM combines the interpretability of discrete choice theory with the flexibility of neural networks. Therefore, it is capable of modeling complex, non-linear interactions while preserving transparency about the influence of individual features. Utilizing features extracted from both cases and judges, the model predicts whether the Supreme Court will reverse a lower court's ruling and highlights the factors most strongly associated with disagreement. When tested on a dataset of Supreme Court opinions, the LLM achieves 80% accuracy in predicting outcomes, outperforming conventional logit and deep learning-based models. Despite the possibility of motivated reasoning in Supreme Court opinions, limiting causal interpretation, the findings show that the LLM presents an interpretable and effective predictive framework applicable to the study of judicial decision-making.