In Ban and Rudin’s (2018) “The Big Data Newsvendor: Practical Insights from Machine Learning,” the authors take an innovative machine-learning approach to a classic problem solved by almost every company, every day, for inventory management. By allowing companies to use large amounts of data to predict the correct answers to decisions directly, they avoid intermediate questions, such as “how many customers will we get tomorrow?” and instead can tell the company how much inventory to stock for these customers. This has implications for almost all other decision-making problems considered in operations research, which has traditionally considered data estimation separately from the decision optimization. Their proposed methods are shown to work both analytically and empirically with the latter explored in a hospital nurse staffing example in which the best one-step, feature-based newsvendor algorithm (the kernel-weights optimization method) is shown to beat the best-practice benchmark by 24% in the out-of-sample cost at a fraction of the speed.

We consider a seller who can dynamically adjust the price of a product at the individual customer level, by utilizing information about customers’ characteristics encoded as a d-dimensional feature vector. We assume a personalized demand model, parameters of which depend on s out of the d features. The seller initially does not know the relationship between the customer features and the product demand but learns this through sales observations over a selling horizon of T periods. We prove that the seller’s expected regret, that is, the revenue loss against a clairvoyant who knows the underlying demand relationship, is at least of order [Formula: see text] under any admissible policy. We then design a near-optimal pricing policy for a semiclairvoyant seller (who knows which s of the d features are in the demand model) who achieves an expected regret of order [Formula: see text]. We extend this policy to a more realistic setting, where the seller does not know the true demand predictors, and show that this policy has an expected regret of order [Formula: see text], which is also near-optimal. Finally, we test our theory on simulated data and on a data set from an online auto loan company in the United States. On both data sets, our experimentation-based pricing policy is superior to intuitive and/or widely-practiced customized pricing methods, such as myopic pricing and segment-then-optimize policies. Furthermore, our policy improves upon the loan company’s historical pricing decisions by 47% in expected revenue over a six-month period. This paper was accepted by Noah Gans, stochastic models and simulation.

The portfolio optimization model has limited impact in practice due to estimation issues when applied with real data. To address this, we adapt two machine learning methods, regularization and cross-validation, for portfolio optimization. First, we introduce performance-based regularization (PBR), where the idea is to constrain the sample variances of the estimated portfolio risk and return, which steers the solution towards one associated with less estimation error in the performance. We consider PBR for both mean-variance and mean-CVaR problems. For the mean-variance problem, PBR introduces a quartic polynomial constraint, for which we make two convex approximations: one based on rank-1 approximation and another based on a convex quadratic approximation. The rank-1 approximation PBR adds a bias to the optimal allocation, and the convex quadratic approximation PBR shrinks the sample covariance matrix. For the mean-CVaR problem, the PBR model is a combinatorial optimization problem, but we prove its convex relaxation, a QCQP, is essentially tight. We show that the PBR models can be cast as robust optimization problems with novel uncertainty sets and establish asymptotic optimality of both Sample Average Approximation (SAA) and PBR solutions and the corresponding efficient frontiers. To calibrate the right hand sides of the PBR constraints, we develop new, performance-based k-fold cross-validation algorithms. Using these algorithms, we carry out an extensive empirical investigation of PBR against SAA, as well as L1 and L2 regularizations and the equally-weighted portfolio. We find that PBR dominates all other benchmarks for two out of three of Fama-French data sets.

No abstract

Problem definition:We study the practice-motivated problem of dynamically procuring a new, short life-cycle product under demand uncertainty. The firm does not know the demand for the new product but has data on similar products sold in the past, including demand histories and covariate information such as product characteristics.Academic/practical relevance: The dynamic procurement problem has long attracted academic and practitioner interest, and we solve it in an innovative data-driven way with proven theoretical guarantees.This work is also the first to leverage the power of covariate data in solving this problem.Methodology: We propose a new, combined forecasting and optimization algorithm called the Residual Tree method, and analyze its performance via epi-convergence theory and computations. Our method generalizes the classical Scenario Tree method by using covariates to link historical data on similar products to construct demand forecasts for the new product.Results: We prove, under fairly mild conditions, that the Residual Tree method is asymptotically optimal as the size of the data set grows. We also numerically validate the method for problem instances derived using data from the global fashion retailer Zara. We find that ignoring covariate information leads to systematic bias in the optimal solution, translating to a 6-15% increase in the total cost for the problem instances under study. We also find that solutions based on trees using just 2-3 branches per node, which is common in the existing literature, are inadequate, resulting in 30-66% higher total costs compared with our best solution.Managerial implications: The Residual Tree is a new and generalizable approach that uses past data on similar products to manage new product inventories. We also quantify the value of covariate information and of granular demand modeling.for apparel articles may include the procurement cost, retail price, colour, item type (e.g., sweater, t-shirt, etc.), fabric, design style (e.g., sporty, classic), expert predictions of the product's popularity, etc. We refer to covariates tied to product attributes that do not change over time as "static." Our approach can also account for "dynamic" covariates-e.g., lagged demand observations included to naturally account for demand auto-correlations. Importantly, covariates are essential for accounting for product heterogeneity by providing a means for aligning demand observations across different products. This allows us to construct a large number of estimated demand trajectories relevant to the new product, which we can then use in a scenario-based model of forecast updating dynamics.Our method employs a new approach which we call the Residual Tree method, which can be used to solve general multi-stage stochastic programs where there is no information about the