Utility-Weighted Forecasting and Calibration for Risk-Adjusted Decisions under Trading Frictions

Battle Count: 67

Directly addresses the core challenge of quantitative trading: converting probabilistic forecasts into implementable trading decisions under realistic transaction costs, market impact, and binding constraints. The UWC framework provides a principled method to reduce decision loss by 30%+ relative to uncalibrated baselines, with the mechanism identified as reducing constraint-binding frequency from 16% to 5%. The friction operator formalization, walk-forward evaluation discipline, and model-risk governance framework are directly applicable to production trading systems. The paper bridges the gap between statistical forecast quality and realized economic performance in a way that is immediately actionable for quant practitioners.

The UWC calibration layer itself is computationally lightweight (post-processing monotone spline warp solved via SLSQP). However, the full pipeline requires: (1) probabilistic forecasting model (GARCH/deep quantile regression), (2) friction operator estimation (spread/impact proxies from market data), (3) convex portfolio optimization (QP/SOCP via cvxpy/OSQP), (4) nested walk-forward protocol with embargo rules, (5) dependence-aware inference (block bootstrap, HAC), and (6) model-risk monitoring. The non-convex extensions (cardinality, discrete execution) add significant computational burden. The pre-committed protocol and governance requirements add operational complexity. Overall, moderate-to-high complexity requiring expertise in both econometrics and optimization.

The paper provides a full Jupyter Notebook implementing the complete pipeline, data dictionary, column mapping, and replication instructions. All tables and figures are reproducible from the provided evaluation panel. Solver settings (OSQP via cvxpy, SLSQP for calibration) are fixed and documented. Pre-committed analysis plan eliminates adaptive specification search. However, raw market microstructure data (quotes, depth) may be subject to vendor licensing restrictions. A synthetic non-tradable evaluation panel is offered as fallback. The evaluation window is narrow (one month, Dec 2025), limiting external validation.