Abstract
Traditional signal-based trading strategies often rely on a rigid, binary selection of principal portfolios, which can lead to poor diversification and instability across varying market regimes. This paper introduces the Linear Trading Position with Sparse Spectrum (LTPSS), a framework that allows for flexible, continuous spectral energies. By formulating the problem as a multi-norm optimization and solving it with a novel fixed-point algorithm featuring guaranteed linear convergence, LTPSS demonstrates superior risk-adjusted performance across seven global benchmark datasets.
Introduction: The Rigidity of Principal Portfolios
In signal-based trading, investors map current market signals to future asset returns using a prediction matrix. A common technique to manage complexity is focusing on Principal Portfolios (PPs)—the primary eigen-directions of this matrix. However, conventional methods typically employ a "hard selection" process: a PP is either fully included (spectral energy = 1) or entirely excluded (spectral energy = 0). This binary approach often results in a lack of diversification, making strategies vulnerable when market conditions shift away from the few selected features. The authors argue that this rigidity is akin to putting all eggs in a few specific baskets, leaving the portfolio exposed to regime shifts.
Methodology: Flexible Weights and Sparse Spectra
The proposed LTPSS method moves beyond binary selection by assigning each PP a weight (spectral energy) between 0 and 1. This allows the model to capture "weaker" but still valuable market signals that are often discarded by traditional methods. The optimization objective incorporates the Frobenius norm for prediction accuracy, the Nuclear norm to promote a sparse spectrum (low-rank structure), and the Spectral norm to enforce a self-financing constraint, ensuring the strategy requires no external cash injections.
To solve this complex, non-differentiable problem, the authors developed a Krasnosel’skiĭ-Mann (KM) fixed-point algorithm. Unlike standard KM implementations that converge slowly, this algorithm is theoretically proven to achieve a linear convergence rate, making it both mathematically robust and computationally efficient for real-time trading applications.
Results and Performance Analysis
LTPSS was rigorously tested against seven diverse benchmarks, including US equity factors (Fama-French), global indices (MSCI), and blue-chip stocks (Stoxx50, FTSE100). These datasets provide a rigorous test of how the model handles different geographical regions and financial climates.
Key findings include:
- Superior Returns: LTPSS consistently achieved the highest Mean Returns (MR). On the FTSE100, it outperformed the standard LTP-PP method by a factor of 25.
- Risk-Adjusted Stability: The strategy showed significantly higher Sharpe Ratios (SR), particularly in international markets like the Stoxx50 and Fund of Funds (FOF) data, where traditional PP methods often struggled or yielded negative results.
- Global Robustness: While baselines like LTP-PP performed well on US-centric data, their performance often degraded on international benchmarks. LTPSS maintained stable and superior performance across all tested regions, proving its adaptability to different market structures.
Conclusion
LTPSS demonstrates that diversification in signal-based trading is not just about the number of assets, but about how spectral energy is distributed across the prediction matrix. By replacing rigid selection with a sparse, continuous spectrum, the model achieves greater adaptability and robustness. While the current framework is linear and highly interpretable, future work may explore non-linear trading positions and integration with broader financial factor libraries to further enhance signal processing power.
Summary updated