Revisiting the Structure of Trend Premia: When Diversification Hides Redundancy

Battle Count: 80

Highly relevant to quantitative trading, particularly for CTA/trend-following strategy design. The paper directly challenges the conventional practice of equal-weighting multiple trend horizons and provides actionable evidence that a barbell (short + long) structure outperforms five-horizon equal weighting. The Bayesian dynamic allocation framework with persistence filtering is directly implementable for systematic trading. The ablation study methodology provides a template for evaluating marginal contribution of signal components. The crisis alpha analysis (Table 11) is relevant for risk management. The paper bridges academic theory (minimum-variance portfolio of correlated factors) with practitioner implementation (CTA replication with realistic transaction costs).

Moderate to high complexity. Requires: (1) Bayesian graphical model with Gaussian state-space dynamics for asset-level exposure estimation; (2) rolling optimization across multiple horizons with stability diagnostics; (3) persistence filtering with percentile-based adaptive thresholds; (4) exponential moving average smoothing; (5) multi-asset universe management with realistic transaction/roll costs. The theoretical framework is elegant but the full pipeline involves multiple interacting components. The algorithm is well-specified in Appendix B, reducing ambiguity. Key challenges include proper Bayesian inference implementation, handling of non-stationary covariance structures, and calibration of persistence thresholds.

The paper provides detailed algorithmic descriptions (Algorithm 1 in Appendix B), mathematical formulations, and parameter specifications (training windows, subwindow lengths, percentile thresholds). However, no code repository is provided, and the specific data sources for futures contracts are not fully detailed. The Bayesian graphical model implementation details reference prior work (Ohana et al., 2022; Benhamou et al., 2024). Cost assumptions are specified in Table 1. Reproduction would require access to historical futures data and implementation of the Bayesian state-space model.