This paper introduces Fused Encoder Networks (FEN), a novel transfer ranking model designed to address data scarcity in cross-sectional momentum strategies, particularly for cryptocurrencies with short trading histories. The model employs a hybrid parameter-sharing architecture: a source encoder block pre-trained on a larger foreign exchange (FX) dataset and a target encoder block trained on the smaller cryptocurrency dataset, with their outputs concatenated and passed through a fully connected layer. The self-attention mechanism within each encoder captures interactions among assets, enhancing ranking performance. The training process involves three steps: pre-training the source model on FX data, training the target model with the source encoder frozen, and fine-tuning the entire network with a low learning rate. Negative transfer is mitigated by selecting source models with high Sharpe ratios on a validation year. Experiments on the top 10 cryptocurrencies (2018-2021) show that FEN significantly outperforms benchmarks including risk-adjusted momentum, MLP, ListNet, LambdaMART, and a self-attention ranker with parameter sharing. FEN achieves a Sharpe ratio of 1.02, a threefold improvement over risk-adjusted momentum and ~50% over the best baseline, with superior ranking accuracy (NDCG@2=0.89) and robustness to transaction costs. The results demonstrate the effectiveness of transfer ranking in finance, offering a practical solution for deploying deep learning strategies on instruments with limited data.