- The paper introduces FlashFolio, a GPU-accelerated convex solver that significantly outperforms CPU-based methods like MOSEK in large-scale portfolio optimization.
- It models single and multi-period portfolio problems with factor risk, linear trading friction, and nonlinear market impact under realistic trading conditions.
- Empirical results show robust performance with speedups up to 48x, enabling latency-sensitive asset management and real-time rebalancing applications.
FlashFolio: GPU-Accelerated Portfolio Optimization for Practical Large-Scale Asset Allocation
FlashFolio addresses the persistent computational bottleneck in large-scale portfolio optimization, particularly relevant for modern asset management and electronic trading operations where asset universes can be several thousand securities and timely rebalancing is imperative. The framework supports both single-period and multi-period portfolio optimization, incorporating factor-based risk models, linear bid–offer spread costs, and nonlinear market impact—each key for realistic institutional portfolio construction.
The single-period model specifies the allocation problem as a regularized risk-constrained optimization with linear trading friction and nonlinear market impact, under exposure and budget constraints. The multi-period formulation generalizes this to joint optimization over a trading trajectory, explicitly modeling dynamic rebalancing paths, path-dependent costs, and time-varying liquidity conditions. Key state variables—portfolio weights and trade increments—are tightly coupled through constraints and nonlinear cost penalties, reflecting the temporal execution and adversarial market microstructure observed in practice.
Methodology: GPU Acceleration and Solver Construction
FlashFolio leverages the architectural features of modern GPUs—massively parallel floating-point units and high memory bandwidth—by recasting the portfolio optimization as a convex program suitable for efficient mapping to GPU primitives using JAX. Critically, the solver design exploits the prevailing block structure and sparsity of factor-model-based risk constraints and the separable nature of trading cost penalties. The pipeline circumvents CPU-centric bottlenecks inherent in branch-and-bound or conic reformulation-based solvers (e.g., MOSEK), utilizing fixed-point iteration with high-precision residual control.
The empirical workflow builds realistic benchmark instances via data-driven generation of alphas, factor risk models, and transaction cost coefficients. Alphas are constructed as noisy forward returns; risks employ multi-factor structures calibrated on rolling historical windows with Fama–French factors; and transaction costs are parameterized by CRSP-derived spreads and a nonlinear ADV-scaled market impact function. For multi-period experiments, liquidity conditions are scheduled following empirically observed intraday U-shapes.
Experimental Evaluation
FlashFolio is systematically benchmarked against MOSEK—an industry-standard conic solver—across hundreds of instances spanning wide ranges of regularization and cost parameters. All runs utilize high-end hardware: NVIDIA H100 for FlashFolio (GPU), and dual Intel Xeon Platinum for MOSEK with equal wall-clock budgeting. Performance is quantified by solved instance counts and shifted geometric mean (SGM1) of runtime.
In all single-period cases, FlashFolio is strictly faster than MOSEK, with speedups ranging from 2.4× up to 12.9×. For multi-period optimization, the acceleration is even more pronounced, scaling up to 48×, while MOSEK fails to solve certain challenging cases that FlashFolio completes robustly. Notably, robustness degrades for MOSEK in regimes with high cost coefficients, whereas FlashFolio remains stable, reliably reaching high-precision fixed points.
Implications and Future Directions
The demonstrated runtime and robustness advantages of GPU-accelerated convex solvers have several practical implications:
- Production-Quality Latency: For asset managers executing systematic strategies under tight intraday or real-time constraints, FlashFolio enables significantly reduced operational latency, facilitating more responsive risk and execution management, particularly in high-dimensional multi-period settings.
- Model Expressivity: By removing computational bottlenecks, richer formulations—including higher granularity of time steps, asset universes, or additional path-dependent market frictions—become practical for live trading or pre-trade analytics.
- Resource Efficiency and Scaling: The results suggest that high-performance optimization previously reserved for CPU clusters can be efficiently migrated to single high-throughput GPU systems, which is economically advantageous and environmentally favorable.
From a theoretical perspective, FlashFolio’s architecture substantiates the utility of direct GPU-based approaches over conic reformulation pipelines for a broad class of convex dynamic optimization problems. The modular design also opens the door to rapid prototyping for extensions such as stochastic programs or integration with reinforcement learning for direct policy calibration.
Future developments could generalize this approach to non-convex settings (e.g., regime-switching models), hybridize with hardware-specific solvers (TPUs or ASICs), or embed differentiable optimization layers within end-to-end deep learning frameworks for market modeling and trade execution.
Conclusion
FlashFolio demonstrates that direct GPU-accelerated convex optimization is not only computationally superior for large-scale factor-based portfolio construction, but also enhances solver robustness in realistic, high-dimensional settings with nonlinear costs. By outperforming CPU-based baselines such as MOSEK, especially for time-coupled multi-period problems, FlashFolio provides a practical pathway towards high-throughput, latency-sensitive portfolio optimization in contemporary financial markets. The approach redefines the feasible regime for deploying complex asset allocation strategies and motivates further exploration of hardware-aligned optimization methods in financial engineering.
