Static chess position classification enables fast evaluation in analysis pipelines, tutoring systems, and other low-latency applications that cannot rely on deep search. We present a comparative study of three deep learning approaches for static position classification, using a uniform and compact 8x8x18 board encoding derived from Forsyth–Edwards Notation (FEN) data. The dataset comprises 1,530,000 chess positions with balanced distribution: 510,000 White Advantage, 510,000 Balanced, and 510,000 Black Advantage positions, sourced from a comprehensive Kaggle chess evaluations dataset containing Stockfish evaluations at depth 22. The task assigns each position to one of three strategic outcomes based on engine evaluations using ±150 centipawn thresholds. Models are trained and tested with 900,000 training, 180,000 validation, and 450,000 test positions using random sampling with seed 42. Performance is measured using accuracy and macro-averaged precision, recall, and F1score, with detailed per-class and confusion matrix analysis. Among the evaluated models, a Custom ResCNN delivers the strongest results, outperforming an Adapted ResNet-12 and a Baseline CapsNet. The benchmark highlights trade-offs between model families under compact inputs and offers a reference point for future research on richer feature sets and architectural refinements.