Kolmogorov–Arnold Networks for Daily Apparel Retail Demand Forecasting: A Comparative Evaluation Across Forecasting Models

Operational apparel forecasting requires comparisons that combine strong classical baselines, modern learning architectures, sensitivity analysis, and evaluation across multiple forecast origins. This study benchmarks nine methods for one-step-ahead daily forecasting using 1403 observations of a retailer-defined potential-sales measure from 11 apparel categories in Türkiye. The methods are seasonal naïve (SNaive), exponential smoothing (ETS), Light Gradient Boosting Machine (LightGBM), feedforward neural network (FFNN), standard Kolmogorov-Arnold Network(KAN), long short-term memory (LSTM), Transformer, TimeKAN, and iTransformer, all evaluated under a lag-only information setting. Evaluation combines a 183-day test window, a five-method joint-versus-per-category ablation, fixed-configuration look-back sensitivity, and seven rolling-origin windows. ETS achieved the lowest macro weighted absolute percentage error (WAPE; 12.02%) and mean absolute scaled error (MASE; 0.690) on the test window, while iTransformer and TimeKAN were the strongest learning-based alternatives. Across rolling origins, iTransformer achieved the lowest mean macro WAPE, whereas ETS led five of seven windows and achieved the lowest mean pooled WAPE and macro MASE. Standard KAN showed intermediate performance, while TimeKAN and iTransformer improved with longer look-back windows under fixed configurations. Per-category configurations yielded lower WAPE and MASE values than matched joint configurations in the five-method ablation. Overall, rankings depended on forecast origin, error aggregation, historical context, and information-sharing design.