Learning to Spend: Model Predictive Control for Budgeting under Non-Stationary Returns

Battle Count: 50

While the paper focuses on marketing budget allocation rather than trading, the core framework is conceptually transferable to quantitative trading. The distinction between predictable structure (exploitable via MPC/forecasting) and unpredictable drift (where reactive strategies suffice) directly parallels the challenge of distinguishing alpha signals from noise in trading. The receding-horizon optimization under constraints, execution noise modeling, and the finding that non-stationarity alone does not justify predictive control are relevant to portfolio rebalancing, algorithmic execution, and market-making strategies. However, the paper does not address financial markets, asset pricing, or trading-specific challenges.

The MPC framework involves constrained nonlinear optimization (solved via IPOPT/CasADi), particle filtering for adaptive tracking, SARIMAX for seasonal forecasting, and a closed-loop simulation environment. The optimization problem is small (12-week horizon, weekly decisions), making computation tractable. However, correctly implementing the particle filter, parameter estimation pipeline, and ensuring proper separation of historical/evaluation data requires careful engineering. The simulation framework itself is moderately complex with multiple interacting components.

The paper provides detailed simulation specifications including budget construction, execution layer dynamics, response curves, and operating regimes. However, no code repository is mentioned, and the synthetic environment parameters (e.g., specific noise levels, budget magnitudes, response curve parameters) are not fully enumerated in the main text. The methodology is well-described but full reproduction would require careful reimplementation of the simulation framework.