Benchmarking Deep Time Series Models for Equity Portfolios

Battle Count: 105

Directly addresses model selection for daily equity portfolio construction. Provides a portable benchmark framework for evaluating forecasting architectures under realistic implementation constraints (transaction costs, capacity, beta, industry, risk, leverage, turnover). The key finding that TS-RIDGE (strong in raw decile sorting) should not be deployed at daily frequency under constrained QP, while TransEnc-8 is the least-costly constrained implementation, is directly actionable for quantitative trading desks. The deployment-adjusted acceptability index provides a principled way to incorporate implementation frictions into model selection.

High complexity due to: (1) 15 model architectures requiring separate training pipelines with walk-forward protocol, (2) SMAA layer with 10,000 moving-block resamples and 10,000 Dirichlet draws over 7-criterion simplex, (3) constrained QP portfolio optimization with multiple constraints (dollar neutrality, leverage, capacity, beta, industry, turnover) solved daily via OSQP, (4) deployment-adjusted acceptability index with entropic regularization, (5) multiple inference layers (bootstrap, MCS, factor regressions), (6) feature engineering with 24 predictors across 3 nested universes, (7) capacity frontier and cost sensitivity analyses. The SMAA complexity is O(BQ|M|K) and the portfolio layer requires repeated sparse first-order QP solves.

The paper provides detailed implementation settings including fixed software environment (Python 3.12.7, PyTorch 2.8.0, OSQP 1.1.1, etc.), fixed random seeds, walk-forward training protocol, feature definitions, model architectures with parameter counts, and portfolio-layer parameters. However, CRSP data requires WRDS license and is not redistributed. The replication scripts and registry keys are documented in Appendix E. The SMAA design uses 10,000 resamples and 10,000 Dirichlet draws with specified seeds.