Dual-objective autoencoder framework for Taiwan 50 index sparse portfolio

This research develops a dual-objective framework for sparse portfolio construction using autoencoders on the Taiwan 50 (0050 TW) index. Standard (AE) and variational (VAE) autoencoders are augmented with dedicated selection and weight layers that jointly learn stock inclusion and allocation from a shared latent representation. The main objectives are to minimize tracking error (TE) and maximize excess return (ER), subject to enforcing cardinality to select approximately 16 % of the constituents from 0050 TW. Empirical evaluation across rolling out-of-sample windows reveals that AE consistently outperforms VAE in ER across all TE weights and exhibits substantially lower downside risk via VaR and CVaR. Adding Lasso penalties to create AE-L and VAE-L effectively reduces TE and reconstruction error through stronger internal latent sparsity. Comparisons with a non-learning MILP-MAD reveal that autoencoder approaches achieve competitive benchmark tracking while offering better ER, particularly at low TE weights. Overall, compact end-to-end learned portfolios effectively replicate and potentially outperform the benchmark.