A data-driven approach for a class of stochastic dynamic optimization problems.

Area of Science:

• Operations Research
• Machine Learning
• Data Science

Background:

• Dynamic stochastic optimization models are crucial for sequential decision-making.
• Existing models often overlook estimation errors and model misspecification in predictive methods.
• Bridging the gap between data and decisions requires integrating predictive and prescriptive analytics.

Purpose of the Study:

• To propose a data-driven prescriptive analytics framework for dynamic decision problems.
• To integrate machine learning (predictive) and dynamic optimization (prescriptive) methods.
• To address estimation errors and model misspecification in decision-making processes.

Main Methods:

• Utilized a Hidden Markov Model (HMM) for uncertainty representation and prediction.
• Employed a distributionally robust dynamic optimization model to account for estimation errors.
• Developed an evaluation framework for assessing out-of-sample performance in rolling horizon schemes.

Main Results:

• Demonstrated superior out-of-sample performance in a dynamic asset allocation case study.
• Showcased the framework's ability to extract valuable information from data for robust decisions.
• Provided an empirical certificate of out-of-sample performance evaluation.

Conclusions:

• The proposed framework effectively integrates machine learning and dynamic optimization.
• It enables robust decision-making by addressing estimation errors and model uncertainty.
• The approach has practical importance and applicability in real-world dynamic decision problems.