EconBase
← Back to paper

The Nonstationary Newsvendor with (and without) Predictions

Extracted main text — title through conclusion, appendix excluded. This is what our citation measures are computed over, published so the extraction can be checked by eye.

99,857 characters · 19 sections · 76 citation commands

Rendered from LaTeX for readability, not typeset faithfully. Citation keys are highlighted; maths is left as source; figures, tables and equation environments are summarised rather than reproduced; unrecognised commands are greyed out so nothing is silently dropped. Email addresses are removed.
center[center omitted — 367 chars of source]
abstract\\ \noindentProblem definition: The classic newsvendor model yields an optimal decision for a “newsvendor” selecting a quantity of inventory, under the assumption that the demand is drawn from a known distribution. Motivated by applications such as cloud provisioning and staffing, we consider a setting in which newsvendor-type decisions must be made sequentially, in the face of demand drawn from a stochastic process that is both unknown and nonstationary. All prior work on this problem either (a) assumes that the level of nonstationarity is known, or (b) imposes additional statistical assumptions that enable accurate {\em predictions} of the unknown demand. Our research tackles the Nonstationary Newsvendor without these assumptions, both with and without predictions. \noindentMethodology/results: We first, in the setting without predictions, design a policy which we prove (via matching upper and lower bounds) achieves order-optimal regret -- ours is the first policy to accomplish this without being given the level of nonstationarity of the underlying demand. We then, for the first time, introduce a model for generic (i.e. with no statistical assumptions) predictions with arbitrary accuracy, and propose a policy that incorporates these predictions without being given their accuracy. We upper bound the regret of this policy, and show that it matches the best achievable regret had the accuracy of the predictions been known. \noindentManagerial implications: Our findings provide valuable insights on inventory management. Managers can make more informed and effective decisions in dynamic environments, reducing costs and enhancing service levels despite uncertain demand patterns. This study advances understanding of sequential decision-making under uncertainty, offering robust methodologies for practical applications with nonstationary demand. We empirically validate our new policy with experiments based on {three} real-world datasets containing thousands of time-series, showing that it succeeds in closing approximately 74% of the gap between the best approaches based on nonstationarity and predictions alone. \\ This paper is published at https://pubsonline.informs.org/doi/10.1287/msom.2024.1168. \\ \\ \textit{Key words:} newsvendor model; decision-making with predictions; regret analysis

Introduction

The newsvendor problem is a century-old model (edgeworth1888mathematical) that remains fundamental to the practice of operations management. In its original instantiation, a “newsvendor” is tasked with selecting a quantity of inventory before observing the demand for that inventory, with the demand itself randomly drawn from a {\em known} distribution. The newsvendor incurs a per-unit underage cost for unmet demand, and a per-unit overage cost for unsold inventory. The objective is to minimize the total expected cost, and the classic result is that the optimal inventory level is a certain problem-specific quantile (depending only on the underage and overage costs) of the demand distribution.

This paper is concerned with a modern instantiation of the same model, consisting of a {\em sequence} of newsvendor problems over time, each with {\em unknown} demand distributions that {\em vary} over time. While this version of the problem is arguably ubiquitous in practice today, it may be worth highlighting a few motivating examples:

itemize• {\bf Cloud Provisioning:} Consider a website which provisions computational resources from a commercial cloud provider to serve its web requests. Such provisioning is typically done {\em dynamically}, say on an hourly basis, with the aim of satisfying incoming requests at a sufficiently high service level. Thus, the website faces a single newsvendor problem every hour, with an hourly “demand” that can (and does) vary drastically over time. • {\bf Staffing:} A more traditional example is staffing, say for a brick-and-mortar retailer, a call center, or an emergency room. Each day (or even each shift) requires a separate newsvendor problem to be solved, with demand that is highly nonstationary.

Despite its ubiquity, this problem is far from resolved, precisely because the demand (or sequence of demand distributions) is both {\em nonstationary} and {\em unknown} -- indeed, the repeated newsvendor with stationary, but unknown demand was solved by levi2015data, and the same setting with known, but nonstationary demand can be treated simply as a sequence of completely separate newsvendor problems. At present, there are by and large two existing approaches to this problem:

enumerate• {\bf Limited Nonstationarity:} One approach is to design policies which “succeed” under limited nonstationarity, i.e. the cost incurred by the policy should be parameterized by some carefully-chosen measure of nonstationarity (e.g. quadratic variation), and nothing else. This approach has proved fruitful across a diverse set of problems ranging from dynamic pricing (keskin2017chasing) to multi-armed bandit problems (besbes2014stochastic) to stochastic optimization (besbes2015non). Most relevant here, the recent work of keskin2021nonstationary applies this lens to the newsvendor setting (we will discuss this work in detail momentarily). This approach yields policies with theoretical guarantees that are quite robust -- no assumption on the demand (beyond the limited nonstationarity) is required. However, this is far removed from practice, where the next approach is more common. • {\bf Predictions:} The second approach is to utilize some sort of {\em predictions} of the unknown demand. These predictions can be generated from simple forecasting algorithms for univariate time-series, all the way to state-of-the-art machine learning models that leverage multiple time-series and additional feature information. Therefore these predictions may contain much more information than past demand data points, such as various features/contexts, or even black-box type information that is non-identifiable. In addition to being the de facto approach in practice, the use of predictions in newsvendor-type problems is well-studied, and in fact provable guarantees exist for many specific prediction-based approaches (ban2019big,huber2019data,oroojlooyjadid2020applying,zhang2023optimal). All such guarantees rely on (at the very least) the demand and potential features being generated from a known family of stochastic models, so that the framework and tools of statistical learning theory can be applied. Absent these statistical assumptions, it is unclear a priori whether the resulting predictions will be sufficiently accurate to outperform robust policies such as those generated in the previous approach. As a concrete example of this, see (ref), which demonstrates on a real set of retail data that prediction accuracy may vary drastically and unexpectedly, even when those predictions are generated according to the same procedure and applied during the same time period.
figure[figure omitted — 508 chars of source]

To summarize, the repeated newsvendor with unknown, nonstationary demand (which from here on we refer to as the {\em Nonstationary Newsvendor}) admits policies with nontrivial guarantees, which can be made significantly better or worse by following predictions. This suggests the opportunity to design a policy that uses predictions {\em optimally}, in the sense that the predictions are utilized when accurate, and ignored when inaccurate. Ideally, such a policy would run without knowledge of (a) the accuracy of the predictions and (b) the method with which they are generated. {\em This is precisely what we accomplish in this paper.}

The Nonstationary Newsvendor, with and without Predictions

The primary purpose of this paper is to develop a policy that optimally incorporates {\em predictions} (defined in the most generic sense possible) into the {\em Nonstationary Newsvendor} problem. Naturally, a prerequisite to this is a fully-solved model of the Nonstationary Newsvendor without predictions. At present this prerequisite is only partially satisfied (via the work of keskin2021nonstationary), so a nontrivial portion of our contributions will be to fully solve this problem.

Without predictions, the Nonstationary Newsvendor consists of a sequence of newsvendor problems indexed by periods $t \in 1,\ldots,T$, each with unknown demand distribution $D_t$. The level of nonstationarity is characterized via a {\em variation parameter} $v\in[0,1]$, where $v=0$ essentially amounts to stationary demand, and $v=1$ is effectively arbitrary (in a little more detail: a deterministic analogue of quadratic variation is applied to the sequence of means $\{\mathbb{E}[D_1],\ldots,\mathbb{E}[D_T]\}$, and $v\in[0,1]$ is the exponent such that this quantity equals $T^v$). Finally, we measure the performance of any policy using {\em regret}, which is the expected difference in the total cost incurred by the policy versus that of an optimal policy that “knows” the demand distributions. At minimum we aim to design a policy that achieves {\em sub-linear} (i.e. $o(T)$) regret, as such a policy would incur a per-period cost that is on average no worse than the optimal, as $T$ grows. We will in fact design policies which achieve {\em order-optimal} regret with respect to the variation parameter $v$.

To this base problem, we introduce the notion of predictions. In each period we receive a prediction $a_t$ of the mean demand $\mu_t = \mathbb{E}[D_t]$ before selecting the order quantity. Our predictions are generic: no assumption is made on how they are generated. We measure the accuracy of the predictions through an {\em accuracy parameter} $a\in[0,1]$, defined such that $\sum_{t=1}^{T} |a_t-\mu_t| = T^a$. Notice that when $a=0$ the predictions are almost perfect, and when $a=1$ the predictions are effectively useless. We will characterize a precise threshold on $a$ (which depends on $v$) that determines when the predictions should be utilized. Our primary challenge will be to design a policy that makes use of the predictions only when they are sufficiently accurate, and {\em without} having access to $a$. As to the variation parameter $v$, we will separately consider policies which do and do not have access to $v$ -- this distinction will turn out to be {\em the} critical factor in classifying what is and is not achievable.

Our Contributions

Our primary contributions can be summarized as follows.

\paragraph{ 1. Nonstationary Newsvendor (without predictions):} We completely solve the Nonstationary Newsvendor problem. This consists of first constructing a policy and proving an upper bound on its regret:

theorem[Informal] There exists a policy which achieves $\tilde{O}(T^{(3+v)/4})$ regret\footnote{The $\tilde{O}(\cdot)$ notation hides logarithmic factors.} without knowing $v$.

We then show that this regret is minimax optimal up to logarithmic factors:

proposition[Informal] No policy can achieve regret better than $O(T^{(3+v)/4})$, even if $v$ is known.

As alluded to earlier, keskin2021nonstationary previously initiated the study of the Nonstationary Newsvendor. Our results are distinct in terms of both modeling and theoretical contributions. We will expound these distinctions more carefully later on.

itemize• {\bf Modeling: } The most crucial difference in our model is that we allow both the demand and the set of possible ordering quantities to be {\em discrete}. This is certainly of practical concern (e.g. physical inventory, employees, and virtual machines are all indivisible units of demand), but moreover we will show that the results of keskin2021nonstationary {\em require} both the demand and set of feasible ordering quantities to be continuous. Thus, there is no overlap in our theoretical results. • {\bf Results:} keskin2021nonstationary succeed in designing a policy that achieves order-optimal regret, but crucially, their policy requires that the variation parameter $v$ be known. In addition to being concerning from a practical standpoint, this leaves open the theoretical question of what exactly is achievable in settings for which $v$ is unknown. Our results show that the {\em same} regret can be achieved without knowing $v$.

\paragraph{ 2. Nonstationary Newsvendor with Predictions:} We construct a policy that {\em optimally} leverages predictions, i.e. it is robust to unknown prediction accuracy. To be precise, the previous contribution offers a policy that achieves $\tilde{O}(T^{(3+v)/4})$ regret, and predictions yield a simple policy that achieves $O(T^a)$ regret, so we would expect that the best possible regret is the minimum of these two quantities. We show this formally:

proposition[Informal] No policy can achieve regret better than $O(T^{\min\{(3+v)/4, a\}})$, even if $v$ and $a$ are known.

Our main algorithmic contribution is a policy which achieves this lower bound (up to log factors) {\em without} knowing the prediction accuracy:

theorem[Informal] There exists a policy which achieves regret $\tilde{O}(T^{\min\{(3+v)/4, a\}})$, knowing $v$, and without knowing $a$.

\addtocounter{theorem}{-2} Finally, since our policy relies on knowledge of the variation parameter $v$, the remaining question is whether the same regret is achievable if both $v$ and $a$ are unknown. We show that in fact predictions {\em cannot} be incorporated in any meaningful way in this case:

proposition[Informal] If $v$ and $a$ are unknown, then no policy can achieve regret better than $O(T^{\max\{(3+v)/4,a\}})$ for all $v,a \in [0,1]$.

\addtocounter{proposition}{-3}

Our theoretical results are summarized in the (ref). Each entry has a corresponding policy that achieves the stated regret, along with a matching lower bound.

table[table omitted — 563 chars of source]

\paragraph{ 3. Empirical Results:} Finally, we demonstrate the practical value of our model (namely the Nonstationary Newsvendor with Predictions) and our policy via empirical results on three real-world datasets that span our motivating applications above: daily web traffic for Wikipedia.com (of various languages), daily foot traffic across the Rossmann store chain, and daily visitors at a certain Japanese restaurant. These datasets together contain over one thousand individual time-series on which we generate predictions of varying quality, using four different popular forecasting and machine learning algorithms. We apply our policy, and compare its performance against the two most-natural baseline policies: our optimal policy without predictions, and the simple policy which always utilizes the predictions (these correspond to the two “existing approaches” described previously). A snapshot of our results, for the Rossmann stores depicted in (ref), is given in (ref).

table[table omitted — 533 chars of source]

More generally, on any given experimental instance (i.e. a time-series and a set of predictions), the minimum (maximum) of the costs incurred by these two baselines can be viewed as the best (worst) we can hope for. Thus we measure performance in terms of the proportion of the gap between these two costs incurred by our policy, so if this “optimality gap” is close to 0, our policy performs almost as good as the better one of the two baselines. Note that {\em randomly} selecting between the two baseline policies yields an (expected) optimality gap of 0.5. We find that in the Rossmann dataset, the average optimality gap is 0.26 when the predictions are accurate, and 0.28 when the predictions are inaccurate. In the Wikipedia dataset, the average optimality gap is 0.40 when the predictions are accurate, and 0.07 when the predictions are inaccurate. In the Restaurant dataset, the average optimality gap is 0.10 when the predictions are accurate, and 0.39 when the predictions are inaccurate. This demonstrates that our policy performs well, irrespective of the quality of the predictions.

Literature Review

The earliest works on the newsvendor model assume that the demand distribution is fully known arrow1958studies,scarf1960optimality. This assumption has then been relaxed, and we can divide the approaches in which the demand distribution is unknown into parametric and nonparametric ones. One of the most popular parametric approaches is the Bayesian approach, where there is a prior belief in parameters of the demand distribution, and such belief is updated based on observations that are collected over time. scarf1959bayes first applied the Bayesian approach to inventory models, and later this is studied in many works karlin1960dynamic,iglehart1964dynamic,azoury1985bayes,lovejoy1990myopic. liyanage2005practical introduced another parametric approach called operational statistics which, unlike the Bayesian approach, does not assume any prior knowledge on the parameter values. Instead it uses past demand observations to directly estimate the optimal ordering quantity.

Nonparametric approaches have been developed in recent years. The first example of a nonparametric approach is the SAA method, first proposed by kleywegt2002sample and shapiro2003monte. levi2007provably applied SAA to the newsvendor problem by using samples to approximate the optimal ordering quantity, and levi2015data improve significantly upon the bounds of levi2007provably for the same problem. Other non-parametric approaches include stochastic gradient descent algorithms burnetas2000adaptive,kunnumkal2008using,huh2009nonparametric and the concave adaptive value estimation (CAVE) method godfrey2001adaptive,powell2004learning. With the development of machine learning, ban2019big and oroojlooyjadid2020applying propose machine learning/deep learning algorithms using demand features and historical data to solve the newsvendor problem.

All the previous studies consider the newsvendor in a static environment where the demand distribution is the same over time. However, in reality the demand distribution is often nonstationary. There are two common practices to resolve this issue. The first is to model the nonstationarity and utilize past demand observations according to the model. One common way is to model the nonstationarity as a Markov chain. For example, treharne2002adaptive applied this idea to inventory management and aviv2005partially and chen2019coordinating applied this idea to revenue management. Another approach is to bound the nonstationarity via a variation budget, which has been applied to stochastic optimization besbes2015non, dynamic pricing keskin2017chasing, multi-armed bandit besbes2014stochastic, newsvendor problem keskin2021nonstationary, among others. Some of these works are applicable in the sense that our problem can be mapped to their settings (e.g. multi-armed bandit such as besbes2015non,karnin2016multi,luo2018efficient,cheung2022hedging), but these connections do not appear to be fruitful. In particular, the multi-armed bandit papers cited above typically consider a {\em limited-feedback} setting rather than the {\em full-feedback} setting explored in this work. Related to feedback, while our study provides a complete characterization of the regret behavior for the nonstationary newsvendor problem with {\em uncensored} demand, practical applications often involve {\em censored} demand. The nonstationary newsvendor problem under censored demand is an interesting direction for future research.

Beyond the bandit literature, it is worth mentioning recent work on online convex optimization (OCO) with limited nonstationarity. When the level of nonstationarity is {\em known}, the standard first-order OCO algorithms can be modified with carefully chosen restarts and updating rules (besbes2015non,yang2016tracking, chen2019nonstationary). There are also recent works that concern {\em unknown} nonstationarity, such as zhang2018adaptive, baby2019online, bai2022adapting, and huang2023stability. Finally, as mentioned before, keskin2021nonstationary is particularly relevant, so we delay a careful comparison to (ref).

The second common practice is to use predictions on the demand distribution of each time period. Predictions can often be obtained e.g. via machine learning, and a recent line of work looks to help decision-making by incorporating predictions. This framework has been applied to many online optimization problems such as revenue optimization munoz2017revenue, balseiro2022single, caching lykouris2021competitive,rohatgi2020near, online scheduling lattanzi2020online, and the secretary problem dutting2021secretaries. In this paper we will combine the nonstationarity framework and the prediction framework on the newsvendor problem.

Finally, most previous works involving algorithms with predictions analyzed algorithms' performances using competitive analysis (e.g. mahdian2012online,antoniadis2020secretary,balseiro2022single,jin2022online) and obtained optimal consistency-robustness trade-offs, where consistency is an algorithm's competitive ratio when the prediction is accurate, and robustness is the competitive ratio regardless of the prediction's accuracy. However, competitive ratio transfers to a regret bound that is linear in $T$. In contrast, we do regret analysis under this framework and design an algorithm that has near-optimal worst-case regret without knowing the prediction quality. Other papers with regret analyses under the prediction model include munoz2017revenue (revenue optimization in auctions), hao2023leveraging (Thompson sampling), hu2024constrained (constrained online two-stage stochastic optimization), and an2024best (online resource allocation).

Model: The Nonstationary Newsvendor (without Predictions)

We begin this section with a formal description of the {\bf Nonstationary Newsvendor}, along with a comparison to the problem of the same name from keskin2021nonstationary. Consider a sequence of newsvendor problems over $T$ time periods labeled $t=1,\dots, T$. At the beginning of each time period $t$, the decision-maker selects a quantity $q_t\in Q$, where $Q$ is a fixed subset of $\mathbb{R}^+$ bounded above by a quantity we denote as $Q_{\max}$.\footnote{All of our results carry through if $Q$ is allowed to depend on $t$.} Then the period's demand $d_t$ is drawn from an (unknown) demand distribution $D_t$, which depends on the time period $t$. These demand distributions are independent over time. Finally a cost is incurred -- specifically, there is a (known) per-unit underage cost $b_t \in [0,b_{\max}]$ and a (known) per-unit overage cost $h_t\in [0,h_{\max}]$, so that the total cost is equal to $$b_t(d_t-q_t)^{+}+h_t(q_t-d_t)^{+},$$ where $x^{+} = \max \{0, x\}$. The decision-maker observes the realized demand $d_t$,\footnote{The demand is not censored here, as is the case in all of the motivating examples in the introduction. The censored version of our problem is an interesting, but separate subject.} and thus the cost. Note that requiring $q_t \in Q$ does not impose any restriction on {\em modeling}, since $Q$ could simply be selected to be $\mathbb{R}$ (as in much of the literature). In fact, introducing $Q$ allows for modeling important practical concerns such as batched inventory or even simply the integrality of physical items. As we will discuss momentarily, this is a non-trivial concern insofar as theoretical guarantees are concerned.

To complete our description of the Nonstationary Newsvendor, we will need to (a) impose a few assumptions on the demand distributions, and then (b) describe how “nonstationarity” is quantified. These are, respectively, the subjects of the following two subsections.

Demand Distributions

We will assume that the demand distributions come from a known, parameterized family of distributions $\mathcal{D}$:

assumptionEvery demand distribution $D_t$ comes from a family of distributions $\mathcal{D}$ satisfying the following: \begin{itemize} • $\mathcal{D} = \{\mathcal{D}_\mu: \mu \in [\mu_{\min},\mu_{\max}]\}$, that is $\mathcal{D}$ is parameterized by a scalar $\mu$ taking values in some bounded interval. • Each distribution $\mathcal{D}_\mu \in \mathcal{D}$ is sub-Gaussian.\footnote{A random variable $X$ is {\em sub-Gaussian} with {\em sub-Gaussian norm} $\|X\|_{\psi_2}$ if $\mathbb{P}\left(|X| > x \right) \le 2 \exp\left( -x^2/\|X\|_{\psi_2}^2\right)$ for all $x \ge 0.$ For sub-Gaussian variables, we have $\mathbb{E}[|X|] \le 3\|X\|_{\psi_2}$.} \end{itemize}

Assumption (ref) is fairly minimal. Parsing it in reverse: the sub-Gaussianity in part (b) allows for many commonly-used variables, such as the Gaussian distribution and any bounded random variable, while letting us eventually apply Hoeffding-type concentration bounds. Part (a) is particularly minimal at the moment, as $\mu$ represents an arbitrary parameterization of $\mathcal{D}$, but will become meaningful when combined with Assumption (ref). The choice of the symbol “$\mu$” might suggest that $\mu$ represents the mean of $\mathcal{D}_\mu$, and indeed this is what we will assume from here on. But it should be emphasized that our taking $\mu = \mathbb{E}[\mathcal{D}_\mu]$ is strictly for notational convenience (because we will frequently need to refer to the means of these distributions): if $\mu$ were any other parameterization of $\mathcal{D}$, we could simply define a mapping from $\mu$ to the mean values.

Now define $C(\mu,b,h,q)$ to be the expected newsvendor cost when selecting quantity $q \in Q$, given underage/overage costs $b$ and $h$, and demand distribution $\mathcal{D}_\mu$: $$C(\mu,b,h,q) = \mathbb{E}_{d\sim \mathcal{D}_\mu}[b(d-q)^+ + h(q - d)^+].$$ The critical assumption, with respect to the parameterization in Assumption (ref)(a), is that the expected cost is well-behaved as a function of $\mu$:

assumptionFor every $b \in [0,b_{\max}]$, $h \in [0,h_{\max}]$, and $q \in Q$, the function $C(\cdot,b,h,q)$ is Lipschitz on its domain $[\mu_{\min},\mu_{\max}]$, i.e. there exists $\ell \in \mathbb{R}^+$ such that for every $\mu_1,\mu_2 \in [\mu_{\min},\mu_{\max}]$, we have $$|C(\mu_1,b,h,q)-C(\mu_2,b,h,q)|\leq \ell|\mu_1-\mu_2|.$$

Note that in the above description, the Lipschitz constant $\ell$ may depend on $b,h,$ and $q$, but by continuity, there exists a single $\ell$ so that the above holds for all $b,h,q$ simultaneously.

Some useful examples of families $\mathcal{D}$ satisfying Assumptions (ref) and (ref) are the following:\footnote{Unfortunately, (ref) is not guaranteed to hold. For example, for the family of distributions $$\mathcal{D}_\mu \sim

cases\mu, \;\; \mu \in [0,1]\\ \mu + \mathrm{Bernoulli}(0.5) - 0.5, \;\; \mu \in (1,2]

,$$ the function $C(\mu,1,1,1)$ is discontinuous (and thus not Lipschitz) at $\mu = 1$. }

enumerate$\mathcal{D}_\mu \sim \mathcal{N}(\mu,\sigma^2)$, the family of normal distributions with fixed variance $\sigma^2$. In this case, $\ell = O(\sigma(b_{\max} + h_{\max}))$. A relaxation is that the variances may vary (continuously) with $\mu$. • $\mathcal{D}_\mu = \mu + \epsilon$, where $\epsilon$ is any mean-zero, sub-Gaussian variable. • The Poisson distribution is frequently used to model demand (since arrivals are often modeled as a Poisson process). While the Poisson distribution is {\em not} sub-Gaussian, any reasonable truncation satisfies our assumptions. For example, $\mathcal{D}_\mu \sim \min\{\mathrm{Poisson}(\mu),K\mu\}$, for some constant $K$. Here, $K$ can be taken to be large enough so that the truncation happens with small probability (in fact, this probability is $O(e^{-K\mu})$).

To understand the reasoning behind Assumption (ref), consider the problem faced at some time $t$. The optimal choice for the decision-maker here is

equation[equation omitted — 80 chars of source]

where $\mu_t$ is the mean of $D_t$ (i.e. $D_t \sim \mathcal{D}_{\mu_t}$), and $C_t(\mu,q) = C(\mu,b_t,h_t,q)$ to simplify the notation.\footnote{As a sanity check, the classical result for the newsvendor problem (arrow1958studies,scarf1960optimality) states that if $Q=\mathbb{R}$, then $q^*_t$ is the $b_t/(b_t+h_t)$-th quantile of $D_t$.} Since $D_t$ is unknown, it is likely that some $q_t \ne q_t^*$ will ultimately be selected, and we could measure the sub-optimality of this decision (i.e. regret, to be defined soon): $C_t(\mu_t,q_t) - C_t(\mu_t,q_t^*)$. It would be natural then to try to characterize this suboptimality as a function of $|q_t - q_t^*|$, but in fact all of the algorithms we will consider “work” by making an estimate $\hat{\mu}_t$ of $\mu_t$, and then selecting $\hat{q}_t \in \text{argmin}_{q\in Q}C_t(\hat{\mu}_t,q)$. So motivated, the purpose of Assumption (ref) is to allow us to “translate” error in our estimate of $\mu_t$ to (excess) costs. The following structural lemma makes this precise, and will be used throughout the paper.

lemmaFix any $b$ and $h$ (we will suppress them from the notation). For any $D_{\mu_1}, D_{\mu_2}\in \mathcal{D}$, let $q_1^* \in \text{argmin}_{q\in Q} C(\mu_1,q)$ and $q_2^* \in \text{argmin}_{q\in Q} C(\mu_2,q)$. Then we have $$C(\mu_1,q_2^*)-C(\mu_1,q_1^*)\leq 2\ell|\mu_1-\mu_2|.$$

Lemma (ref) states that estimation error of the mean $\mu_t$ translates {\em linearly} to excess cost. The proof of Lemma (ref) appears in Appendix (ref).

\paragraph{ Aside: Comparison to keskin2021nonstationary:} The final component in describing the Nonstationary Newsvendor is defining a proper quantification of nonstationarity. Before doing so, we delineate the {\em modeling} differences between our Nonstationary Newsvendor and that of keskin2021nonstationary. There are two primary differences:\footnote{Other minor differences: keskin2021nonstationary require the demand distribution to be bounded, and this assumption is easily relaxed.}

enumerate• The demand distributions $D_t$ in keskin2021nonstationary are assumed to be of the form $D_t=\mu_t+\epsilon_t$, where $\mu_t$ is the mean of $D_t$ that drifts across time and $\epsilon_t$ is the noise distribution that is i.i.d., continuous, and bounded. Effectively, the demand distribution fall into a {\em non-parametric} family of distributions with the same “shape”. In contrast, our demand distributions fall into a {\em parametric} family of distributions, though not necessarily of the same “shape”. • Our set of allowed order quantities $Q$ is bounded, but otherwise arbitrary. In particular, it need not contain the optimal unconstrained order quantity $\text{argmin}_{q\in \mathbb{R}}C(\mu,q)$ for each $\mu$ (or any $\mu$, for that matter). keskin2021nonstationary assume $Q = \mathbb{R}^+$.\footnote{While not stated explicitly, the results in keskin2021nonstationary only require $Q$ to contain points arbitrarily close to every optimal unconstrained order quantity.}

Besides the practical reasons why discrete quantities arise in practice (non-divisible items, batched inventory, etc.), the primary consequence of either of the two differences above is that they preclude a critical lemma used in keskin2021nonstationary (and in fact by levi2007provably) which states that $C(\mu_1,q_2^*)-C(\mu_1,q_1^*)$ (as defined in our Lemma (ref)) scales as $(q_1^* - q_2^*)^2$. This scaling does {\em not} necessarily hold when either the demand distribution or $Q$ is discrete. These relaxations in assumptions yield different lower bounds in the worst-case regret from keskin2021nonstationary, which we will discuss in detail later.

Demand Variation

Just as in keskin2021nonstationary (and keskin2017chasing before that), we measure the level of nonstationarity via a deterministic analogue of quadratic variation for the sequence of means $\mu_1,\ldots,\mu_T$. Specifically, define a {\em partition} of the time horizon $\{1,\dots, T\}$ to be any subset of time periods $\{t_0,\dots, t_K\}$ where $1\leq t_0<\cdots < t_K\leq T$. Here the subset can have any size between 1 and $T$, i.e. $0\leq K\leq T-1$. Then for any sequence of means $\bm{\mu}=\{\mu_1,\dots,\mu_T\}$, its demand variation is

equation[equation omitted — 183 chars of source]

where $\mathcal{P}$ is the set of all partitions.

To motivate the use of partitions in the definition of $V_{\bm{\mu}}$, it is worth contrasting with a measure that may feel more natural, namely the sum of squared differences ({\em SSD}) between consecutive terms, $\sum_{t=2}^{T}\left(\mu_{t}-\mu_{t-1}\right)^2$, which corresponds to taking the densest possible partition $\{1,2,\ldots,T\}$. The maximum in the definition of $V_{\bm{\mu}}$ is {\em not} necessarily achieved by selecting the densest possible partition, but rather by setting $t_0,\dots, t_K$ to be the periods when the sequence $\mu_1,\dots,\mu_T$ changes direction. Thus, the demand variation penalizes {\em trends}, or consecutive increases/decreases, more so than the SSD. For example, the mean sequences $\bm{\mu_1}=\{1,2,3,4,5\}$ and $\bm{\mu_2}=\{1,0,1,0,1\}$, respective variations $V_{\bm{\mu_1}}=(5-1)^2=16$ and $V_{\bm{\mu_2}}=1^2+1^2+1^2+1^2+1^2=5$, despite having identical SSDs.

All of our theoretical guarantees (upper and lower bounds) will be parameterized by $V_{\bm{\mu}}$. This quantity of course depends on $T$, and so it is natural to allow $V_{\bm{\mu}}$ to grow $T$. It will turn out that the most natural parameterization of this growth is via what we will simply call the {\bf variation parameter} $v\in [0,1]$, such that $V_{\bm{\mu}} = BT^v$, where $B$ is some constant (which we take to be equal to one from here on). We denote the set of demand distribution sequences $\{D_1,\dots D_T\}$ whose means $\bm\mu=\{\mu_1,\dots,\mu_T\}$ satisfy $V_{\bm{\mu}}\leq T^v$ as $$\mathcal{D}(v)=\{\{D_1,\dots D_T\}: D_t\in\mathcal{D}\text{ for all }t\text{ and } V_{\bm{\mu}}\leq T^v\}.$$ In the next section, we will show via a minimax lower bound that non-trivial guarantees are only achievable when $v < 1$, and provide an algorithm which achieves the same bound.

\paragraph{ Aside: Time-Series Modeling:} At this point, we have fully described our model for the Nonstationary Newsvendor. All that remains is to define our performance metric, which we will do in the next subsection. We conclude this subsection with an important practical consideration with respect to time-series models and our variation parameter.

Consider, as an example, the following class of time-series models:

equation[equation omitted — 80 chars of source]

Here, $R(t)$ represents a deterministic (and usually simple, e.g. linear) function representing some notion of “trend,” and $S(t)$ represents a deterministic, periodic function representing some notion of “seasonality.” Finally, all stochastic behavior is captured by the random variables $\epsilon_t$, which are assumed to be independent and mean-zero. This time-series model is classic, and yet drives forecasting algorithms (e.g. exponential smoothing) which are still competitive in modern forecasting competitions (makridakis2000m3).

The above model raises an important practical issue: if there exists any (non-trivial) trend $R(\cdot)$ or seasonality $S(\cdot)$, then the demand variation of the sequence of means $\mu_t = R(t) + S(t)$ would scale at least as $T$, meaning $v = 1$ and no meaningful guarantee will be achievable. Our main observation is that time-series effects like trend and seasonality are easily detected and estimated, so that in any practical setting, estimates $\hat{R}(\cdot)$ and $\hat{S}(\cdot)$ should be available, and used to “de-trend” and “de-seasonalize” the data. Concretely, the Nonstationary Newsvendor would take place on the sequence $$ \tilde{d_t} = d(t) - \hat{R}(t) - \hat{S}(t) = (R(t)-\hat{R}(t)) + (S(t)-\hat{S}(t)) + \epsilon_t. $$ The resulting sequence of means $\mu_t = (R(t)-\hat{R}(t)) + (S(t)-\hat{S}(t))$ does not stem from the trend and seasonality, but rather the error in estimating the trend and seasonality. It is this error that is assumed to be nonstationary, but with reasonable variation parameter.

Performance Metric: Regret

We conclude this section by formally defining our performance metric for any policy. A policy is simply a sequence of mappings $\pi=\{\pi_1,\dots,\pi_T\}$, where each $\pi_t$ is a mapping from $d_1,\dots, d_{t-1}$ to an order quantity $q_t\in Q$ at time $t$ (by convention, $\pi_1$ is a constant function).\footnote{Note that we are not considering randomized policies here, but all of our theoretical results (the lower bounds, in particular) hold even when randomization is allowed.} We measure the performance of a policy by its regret. Fix a sequence of demand distributions $\bm{D} = \{D_1,\dots, D_T$\}. Following the earlier notation from Eq. (ref), the regret incurred by a policy which selects order quantities $q_1,\ldots,q_T$ is $$\mathbb{E}^\pi_{\bm{D}}\left[\sum_{t=1}^T\left(C_t(\mu_t,q_t)-C_t(\mu_t, q^*_t)\right)\right],$$ where the expectation is with respect to the randomness of the realized demands. Recall that the demand distributions are independent, so $q_t^*$ as defined in Eq. (ref) depends only on $D_t$. In words, the regret measures the difference between the (expected) total cost incurred by the policy and that of a clairvoyant that knows the underlying demand distributions $\bm{D} = \{D_1,\dots, D_T$\}.\footnote{Note that this is different from a clairvoyant that knows the realized demands $d_1,\ldots,d_T$. Such a clairvoyant would incur zero cost.}

We will be concerned with the {\bf worst-case regret} of a policy across {families} of instances (i.e. sequences of demand distributions) controlled by the variation parameter $v$: $$\mathcal{R}^\pi(T)=\sup_{\bm{D}\in \mathcal{D}(v)} \mathbb{E}^\pi_{\bm{D}}\left[\sum_{t=1}^T\left(C_t(\mu_t,q_t)-C_t(\mu_t, q^*_t)\right)\right].$$ Note that if the worst-case regret $\mathcal{R}^\pi(T)$ of some policy is sublinear in $T$, then that policy is essentially cost-optimal on average as $T$ goes to infinity. In the next section, we will prove a lower bound on the achievable across all policies, and describe an algorithm which achieves this lower bound.

Solution to the Nonstationary Newsvendor (without Predictions)

This section contains a complete solution (i.e. matching lower and upper bounds on regret) to the Nonstationary Newsvendor. We begin with the lower bound:

proposition[Lower Bound: Nonstationary Newsvendor] For any variation parameter $v\in[0,1]$, and any policy $\pi$ (which may depend on the knowledge of $v$), we have $$\mathcal{R}^{\pi}(T)\geq cT^{(3+v)/4},$$ where $c>0$ is a universal constant.

(ref) is a corollary of a more general lower bound ((ref) in the next section) -- it will turn out the Nonstationary Newsvendor is a special case of the Nonstationary Newsvendor with Predictions -- so the proof is omitted. (ref) states that the regret of any policy is at least $\Omega(T^{(3+v)/4})$. It is useful to contrast this with two existing results:

enumerate• {\bf Stationary Newsvendor:} In the special case of i.i.d. demand, it is known that the optimal achievable regret is $\Theta(T^{1/2})$ -- Example 1 of besbes2013implications demonstrates the lower bound, and the SAA method of levi2007provably,levi2015data achieves the upper bound. This point might appear to be incompatible with our result, which states a lower bound of $\Omega(T^{3/4})$ when $v=0$, but in fact the case of $v=0$ is more general than i.i.d. demand since it allows $O(T^0)=O(1)$ demand variation, while i.i.d. demand amounts to zero demand variation. Indeed, our proof of (ref), for the special case of $v=0$, utilizes instances for which the demand distribution is allowed to change $T^{1/2}$ times (by an amount of $T^{-1/4}$, resulting in $O(1)$ variation). As an aside, this discussion raises a natural question: is the disconnect here between stationary (i.i.d.) demand and variation parameter $v=0$ a consequence of our use {\em quadratic} variation, and would the same disconnect arise for other measures of demand variation? In Appendix (ref), we answer both questions in the affirmative by showing that if the exponent $2$ in the demand variation ((ref)) is instead some ${\theta} \ge 0$, then (ref) generalizes to a lower bound of $\Omega(T^{(1+{\theta}+v)/(2+{\theta})})$. Thus, for any ${\theta} > 0$, the case of variation parameter $v=0$ is meaningfully more general than stationary demand. • {\bf Continuous Newsvendor:} A similar “story” plays out in the setting of keskin2021nonstationary, which recall (among other key differences with our model, as described in Section (ref)) requires the additional assumption that both the demand distributions and the possible order quantities be continuous. keskin2021nonstationary show an optimal achievable regret of $\Theta(T^{(1+v)/2})$, which can be contrasted with the stationary (i.i.d.) setting for which an $\Omega(\log T)$ lower bound exists (besbes2013implications). The following table summarizes these lower bounds: \begin{center} \begin{tabular}[h]{@llcc@} \toprule && {\bf Continuous} & {\bf General} \\ \midrule {\bf Stationary (i.i.d.)} && $\log T$ & $T^{1/2}$ \\ {\bf Nonstationary} && $T^{(1+v)/2}$ & $T^{(3+v)/4}$ \\ \bottomrule \end{tabular} \end{center}

In the next two subsections, we will first analyze a simple algorithm which achieves the lower bound of (ref) when the variation parameter $v$ is known, and then use this as a building block for an algorithm which achieves the same bound when $v$ is unknown.

Upper Bound with Known Variation Parameter $v$

If we assume that $v$ is known, then designing a policy which achieves regret matching (ref) is fairly straightforward. In fact, a simple policy based on averaging a fixed number of past demand observations does the job (keskin2021nonstationary use the same policy). That policy, which we call the {\bf Fixed-Time-Window Policy} is defined in Algorithm (ref).

algorithm[algorithm omitted — 572 chars of source]

The Fixed-Time-Window Policy uses a carefully-selected “window” size $n$ that is on the order of $T^{(1-v)/2}$. At each time period $t$, it constructs an estimate $\hat{\mu}_t$ of the mean by averaging the observed demands from the previous $n$ periods, and then selects the optimal order quantity corresponding to $\hat{\mu}_t$. Note that Algorithm (ref) also includes a “scaling constant” $\kappa$ -- this should be thought of as a practical tuning parameter, but for the coming theoretical result, it can be chosen arbitrarily (e.g. $\kappa=1$ suffices).

The following result bounds the worst-case regret of the Fixed-Time-Window Policy:

lemma[Upper Bound: Nonstationary Newsvendor with Known $v$] Fix any variation parameter $v\in[0,1]$. The Fixed-Time-Window Policy $\pi^{\mathrm{fixed}}$ achieves worst-case regret $$\mathcal{R}^{\pi^{\mathrm{fixed}}}(T)\leq CT^{(3+v)/4},$$ where $C\le 3\max\{b_{\max},h_{\max}\}(\delta + Q_{\max})+\ell(2\sqrt{\kappa}+\sqrt{\frac{\pi}{36e}}\frac{\delta}{\sqrt{\kappa}})$, and $\delta = \sup_{\mathcal{D}_\mu \in \mathcal{D}}\|\mathcal{D}_\mu\|_{\psi_2}$.

As promised, Lemma (ref) shows that the Fixed-Time-Window Policy achieves regret that matches the lower bound in (ref). Its proof can be found in Appendix (ref), and amounts to bounding the estimation error incurred by demand noise (which is worse for smaller time windows) and demand mean variation (which is worse for larger time windows). The exact time window used in the policy comes from balancing these two sources of error.

Upper Bound with Unknown Variation Parameter $v$

The lower bound in (ref) holds for policies that “know” $v$. Naturally, it also holds for policies that do not know $v$, but an unanswered question at the moment is whether (a) the lower bound should be even larger when $v$ is unknown, or (b) there exists a policy that matches (ref) without knowing $v$. We show here that case (b) holds by constructing such a policy.\footnote{As a final comparison to keskin2021nonstationary, they do not consider the unknown $v$ setting.} Our policy, which we call the Shrinking-Time-Window Policy (Algorithm (ref)), at a high level uses the Fixed-Time-Window Policy with the smallest variation parameter that is consistent with the demand observed so far. In more detail:

enumerate• It begins with a discrete set of candidate variation parameters $\mathcal{V} = \{v_1,\ldots,v_k\}$: \begin{align} v_j &= \left(1 + \frac{1}{\log T} \right)^{j-1} \frac{1}{\log T}, \quad j=1,\ldots,k \end{align} where $k$ is chosen so that $v_{k-1} < 1 \le v_k$. $\mathcal{V}$ is defined specifically so that the variation parameters are increasing ($v_{j-1} < v_{j}$), and so that it discretizes the interval $[0,1]$ at a sufficiently fine granularity. • At any time period $t$, there is a “current” candidate parameter $v_i$ (initialized to be $v_1$ at $t=1$) that is assumed to be the true variation $v$, and so the corresponding Fixed-Time-Window Policy is applied: a time window of \begin{align} n_i &= \lceil \kappa T^{(1-v_i) / 2}\rceil \end{align} is used, and an estimate of $\mu_t$ is made: \begin{equation} \hat{\mu}_{t}^i = \frac{1}{n_i} \sum_{s=t-n_i}^{t-1} d_{s}, \quad rounded to the nearest value in [\mu_{\min},\mu_{\max}]. \end{equation} • The index $i$ of the “current” candidate parameter $v_i$ is incremented at any period in which the policy gathers sufficient evidence that $v_i < v$. This is possible due to the following observation: if $v_i\approx v$, then by (ref) we have that for any $v_j>v_i$, the regret incurred by the Fixed-Time-Window Policy corresponding to $v_j$ is $O(T^{(3+v_j)/4})$, and thus the {\em cumulative difference} between the estimated mean demands ($|\hat{\mu}_t^i-\hat{\mu}_t^j|$) cannot exceed $O(T^{(3+v_j)/4})$. Thus if this is observed for some $v_j > v_i$, we can conclude that $v_i < v$, and $i$ is incremented.
algorithm[algorithm omitted — 820 chars of source]

This policy's regret matches (up to log factors) the lower bound in (ref):

theorem[Upper Bound: Nonstationary Newsvendor with Unknown $v$] For any variation parameter $v\in[0,1]$, the Shrinking-Time-Window Policy $\pi^{\mathrm{shrinking}}$ achieves worst-case regret $$\mathcal{R}^{\pi^{\mathrm{shrinking}}}(T)\leq C T^{(3+v)/4} \log^{5/2} T,$$ where $C\le 3\max\{b_{\max},h_{\max}\}(\delta + Q_{\max})+12e^{1/4}\ell\left(\gamma+\sqrt{\kappa}\right)+e^{1/4}C_{\mathrm{\cref{upper bound on regret: past-demand-only}}}$, and $C_{\mathrm{\cref{upper bound on regret: past-demand-only}}}$ is the constant in (ref).

The proof of (ref) can be found in Appendix (ref), but at a high level works as follows:

proof[Proof Sketch of Theorem (ref)] First note that the total regret incurred during the first $T^{3/4}$ time periods is at most $O(T^{3/4})$. After that, the total regret incurred during the time periods in which the {\bf if} condition in Algorithm (ref) is triggered is at most $O(k)$, where $k$ is the number of candidate variation parameters defined in (ref), and $k \approx \log^2 T$. Thus, it suffices to bound the total regret incurred {\em between} successive triggerings of the {\bf if} condition. As a final reduction before proceeding, consider the smallest candidate variation parameter that is at least $v$, i.e. define $\ell$ to be the smallest index such that $v\leq v_{\ell}$. Because $v_\ell=\left(1+\frac{1}{\log T}\right)v_{\ell-1}$, we have that $T^{v_{\ell}}$ is a constant multiple away from $T^v$. Thus, it will suffice to bound the total regret by $O(T^{(3+v_\ell)/4} \log^{5/2} T)$. To do this, we first show that (with high probability) throughout the algorithm, the running index $i$ never exceeds $\ell$. To see this, consider the following steps: \begin{enumerate} • For every $j\geq \ell$, because $v\leq v_j$, by Lemma (ref) the Fixed-Time-Window Policy corresponding to the window size $n_j$ has worst-case regret $O(T^{(3+v_j)/4})$. In addition, we can show with high probability (via Hoeffding’s inequality) that $\sum_{s=n_j+1}^{T}|\hat{\mu}_s^{j}-\mu_s| \le \left(\gamma\sqrt{\log T}+\sqrt{\kappa}\right)T^{(3+v_j)/4}$ for every $j\geq \ell$. We may assume this event occurs from now on. • For the sake of contradiction, suppose $i$ exceeds $\ell$ at some period $t$, or equivalently, there is a period $t$ in which $i \ge \ell$, and the {\bf if} condition is triggered with some $j > i$. Then we have \begin{align*} \sum_{s=t_{if}}^{t}|\hat{\mu}_s^{i}-\hat{\mu}_s^{j}|&\overset{(a)}{\leq}\sum_{s=t_{if}}^{t}|\hat{\mu}_s^{i}-\mu_s|+\sum_{s=t_{if}}^{t}|\hat{\mu}_s^{j}-\mu_s|\\ &\overset{(b)}{\leq}\sum_{s=n_i+1}^{T}|\hat{\mu}_s^{i}-\mu_s|+\sum_{s=n_j+1}^{T}|\hat{\mu}_s^{j}-\mu_s|\\ &\overset{(c)}{\leq}\left(\gamma\sqrt{\log T}+\sqrt{\kappa}\right)\cdot T^{(3+v_i)/4}+\left(\gamma\sqrt{\log T}+\sqrt{\kappa}\right)\cdot T^{(3+v_j)/4}\\ &\overset{(d)}{<}2\left(\gamma\sqrt{\log T}+\sqrt{\kappa}\right)\cdot T^{(3+v_j)/4}, \end{align*} where $(a)$ is the triangle inequality and $(b)$ holds because $t_{\textbf{if}}>T^{3/4}$ and $n_i,n_j\leq \kappa T^{1/2}$; since $j>i\geq \ell$, by our assumption in the previous step we get $(c)$; $(d)$ follows since $v_j>v_i$. But this directly contradicts our assumption that the {\bf if} condition is triggered at period $t$. Therefore $i$ never exceeds $\ell$. \end{enumerate} Now suppose two consecutive if conditions occur at times $t'$ and $t''$. Between these periods, we may apply the negation of the {\bf if} condition for any $j > i$, and since $i$ never exceeds $\ell$, we specifically can take $j = \ell$. This yields $\sum_{s=t'+1}^{t''-1}|\hat{\mu}_s^{i}-\hat{\mu}_s^{\ell}|< 2\left(\gamma\sqrt{\log T}+\sqrt{\kappa}\right) T^{(3+v_\ell)/4}$. Then we have \begin{align*} \sum_{s=t'+1}^{t”-1}|\hat{\mu}_s^{i}-\mu_s|&\overset{(a)}{\leq}\sum_{s=t'+1}^{t”-1}|\hat{\mu}_s^{\ell}-\mu_s|+\sum_{s=t'+1}^{t”-1}|\hat{\mu}_s^{i}-\hat{\mu}_s^{\ell}|\\ &\overset{(b)}{\leq}\sum_{s=n_i+1}^{T}|\hat{\mu}_s^{\ell}-\mu_s|+\sum_{s=t'+1}^{t”-1}|\hat{\mu}_s^{i}-\hat{\mu}_s^{\ell}|\\ &\overset{(c)}{<}\left(\gamma\sqrt{\log T}+\sqrt{\kappa}\right)\cdot T^{(3+v_{\ell})/4}+2\left(\gamma\sqrt{\log T}+\sqrt{\kappa}\right)\cdot T^{(3+v_{\ell})/4}\\ &=3\left(\gamma\sqrt{\log T}+\sqrt{\kappa}\right)\cdot T^{(3+v_\ell)/4}, \end{align*} where $(a)$ is the triangle inequality and $(b)$ holds because $t'>T^{3/4}$ and $n_i\leq \kappa T^{1/2}$; the first part of $(c)$ follows by our high-probability assumption, and the second part of $(c)$ follows since the if condition is not triggered between time $t'$ and time $t''-1$; Therefore between two consecutive if conditions, the worst-case regret incurred is $O(T^{v_{\ell}}\sqrt{\log T})$. Because the \textbf{if} condition can happen at most $k\approx \log^2 T$ times, the total worst-case regret of the Shrinking-Time-Window Policy is $O(T^{v_{\ell}}\log^{5/2} T)=O(T^{v}\log^{5/2} T)$.

This concludes our discussion of the Nonstationary Newsvendor. In the next section, we turn to the second subject of this paper, which is the same problem with predictions.

The Nonstationary Newsvendor with Predictions

As described in the introduction, it is likely that when the Nonstationary Newsvendor is faced practice, some notion of a “prediction” of future demand will be made. Such predictions can come from a diverse set of sources ranging from simple human judgement, to forecasting algorithms built on previous demand data, to more-sophisticated machine learning algorithms trained on feature information. The process of sourcing or constructing such predictions is orthogonal to our work. Instead, we treat these predictions as given to us endogenously (and in particular, we make no assumption on the accuracy of these predictions), and attempt to use these predictions optimally.

Model

The {\bf Nonstationary Newsvendor with Predictions} problem assumes all of the setup, assumptions, and notation of the previous Nonstationary Newsvendor problem. In addition, at each time period $t$, we assume that the decision-maker receives a {\bf prediction} $a_t$ before selecting an order quantity $q_t \in Q$.\footnote{We are taking the predictions to be entirely deterministic, so for example, $a_t$ is not allowed to depend on the previously-observed demands $d_1,\ldots,d_{t-1}$. Our results hold if we extend to the setting in which the predictions are stochastic (and adapted to the demand filtration).} This prediction is meant to be an estimate of $\mu_t$, and so we measure the prediction error of a sequence $\bm{a}=\{a_1,\dots, a_T\}$ with respect to a sequence of means simply as $$\sum_{t=1}^{T}|a_t-\mu_t|.$$ Note that unlike demand variation, we have not used partitions here (and in fact, introducing partitions would not have any effect since we are measuring {\em absolute} rather than squared differences). Intuitively, we do not want to require the sequence of errors to be meaningful time-series: the predictions are generic, and their accuracy is allowed to change rapidly. Just as for the demand variation, the prediction error is expected to grow with the time horizon $T$, and the proper parameterization of this growth is via an exponent: we call the {\bf accuracy parameter} the smallest $a \in [0,1]$ such that the prediction error is at most $T^a$. We will always assume that $a$ is unknown to the decision-maker.

algorithm[algorithm omitted — 324 chars of source]

Naturally, the notion of a policy $\pi$ expands to include the predictions: $\pi=\{\pi_1,\dots,\pi_T\}$, where each $\pi_t$ is a mapping from $d_1,\dots, d_{t-1}$ and $a_1,\ldots,a_t$ to an order quantity $q_t\in Q$. The simplest policy, which “should” be used if the prediction error is known to be sufficiently small, is to simply behave as if the predictions were perfect. We call this the {\bf Prediction Policy} (Algorithm (ref)). The following observation collects a few (likely unsurprising) facts about the performance of this policy, with respect to worst-case regret (generalized in the “obvious” manner to incorporate prediction accuracy via the accuracy parameter $a$):

observation[Upper and Lower Bounds: Prediction Policy] Fix any variation parameter $v\in[0,1]$ and any accuracy parameter $a\in[0,1]$. \begin{itemize} • The Prediction Policy $\pi^{\mathrm{prediction}}$ achieves worst-case regret $$\mathcal{R}^{\pi^{\mathrm{prediction}}}(T)\leq CT^{a},$$ where $C\le 2\ell$. • For any policy $\pi$ (which may depend on the knowledge of $a$) that is solely a function of the predictions (i.e. does not depend on the observed demands), we have $$\mathcal{R}^{\pi}(T)\geq cT^{a},$$ where $c>0$ is a universal constant. \end{itemize}

(ref)a) states that the Prediction Policy translates prediction error directly to regret (incidentally, it does this without “knowing” $a$). There are of course other ways in which the predictions could be used, but (ref)b) essentially states that there is nothing to be gained by doing so (even if $a$ is known). The proof of (ref)a) appears in Appendix (ref). (ref)b) is a direct corollary of (ref), which is given in the next subsection.

Extreme Cases

What exactly is achievable for the Nonstationary Newsvendor with Predictions depends heavily on whether or not $v$ and $a$ are known to the policy. To see this, it is worth first considering the two extremes.

\paragraph{ Case 1: Known $v$ and $a$:} A simple policy is available when $v$ and $a$ are both known. Compare the quantities $(3+v)/4$ and $a$. If the former quantity is smaller, apply the Fixed-Time-Window Policy. If the latter is smaller, apply the Prediction Policy. (ref) together imply that this achieves a worst-case regret of $O(T^{\min\{(3+v)/4,a\}})$. This is optimal, as demonstrated by the following result:

proposition[Lower Bound: Known $v$ and $a$] Fix any variation parameter $v\in[0,1]$ and any accuracy parameter $a\in[0,1]$. For any policy $\pi$ (which may depend on the knowledge of $v$ and $a$), we have $$\mathcal{R}^{\pi}(T)\geq cT^{\min\{(3+v)/4, a\}},$$ where $c>0$ is a universal constant.

The proof of this result can be found in Appendix (ref), and relies on an explicit construction of a family of problem instances. Our construction breaks the total time horizon into cycles wherein the demand distribution is i.i.d.. We tune the length of each cycle to be small enough so that it is (provably) hard to detect the change in demand distributions and the predictions are essentially useless for most time periods in the cycle, and large enough so that the demand variation is within $T^v$ and the prediction error is within $T^a$.

\paragraph{ Case 2: Unknown $v$ and $a$:} At the opposite extreme, if $v$ and $a$ are both unknown, is it still possible to achieve $O(T^{\min\{(3+v)/4,a\}})$ worst-case regret? The answer is no:

proposition[Lower Bound: Unknown $v$ and $a$] For any policy that does not depend on the knowledge of $v$ or $a$, there exists a problem instance such that $a \ne (3+v)/4$, and the policy incurs regret at least $cT^{\max\{(3+v)/4,a\}}$ on the instance, where $c>0$ is a universal constant.

(ref) states that the best we can hope for, when $v$ and $a$ are unknown, is a worst-case regret of at least $\Omega(T^{\max\{(3+v)/4,a\}})$. Indeed, it implies that no algorithm can achieve regret $O(T^{f(v,a)})$ for a function $f:[0,1] \times [0,1] \to [0,1]$ satisfying $f(v,a) \le \max\{(3+v)/4,a\}$ for all $v,a\in [0,1]$ and $f(v,a) < \max\{(3+v)/4,a\}$ for some $v,a\in [0,1]$. Note that (ref) shows {\em there exists} a pair of $v$ and $a$, and a corresponding problem instance, such that this lower bound holds. This is in contrast to a result showing that {\em for any} pair of $v$ and $a$, there exists a problem instance such that the lower bound holds, as is common in the literature (e.g. Theorem 1 of keskin2017chasing). This lower bound is easily achieved, for example by applying the Shrinking-Time-Window Policy or the Prediction Policy (or any blind randomization of the two). The proof of (ref) is in Appendix (ref). In contrast to Case 1, the lower bound construction here relies heavily on the fact that we do not know which one of $(3+v)/4$ and $a$ is smaller.

Final Solution

We have finally reached the problem which motivates this entire paper: designing an optimal policy for the Nonstationary Newsvendor with Predictions when the prediction error $a$ is unknown. We will assume that $v$ is known, since when $v$ is unknown, (ref) rules out the possibility of using the predictions to improve on what is already achievable without predictions. On the other hand, by (ref), the absolute best we could hope for is a policy which achieves a worst-case regret of $O(T^{\min\{(3+v)/4, a\}})$. In words, we would like a policy which, without knowing $a$, achieves the same regret had $a$ been known. Our main result is the design of such a policy.

Our policy is called the Prediction-Error-Robust Policy (PERP), and is given in Algorithm (ref). PERP utilizes the Fixed-Time-Window policy $\pi^{\mathrm{fixed}}$ in Section (ref) as an estimate of the true mean to track the quality of the predictions over time.

algorithm[algorithm omitted — 760 chars of source]
theorem[Upper Bound: Known $v$ and Unknown $a$] For any variation parameter $v\in[0,1]$ and any accuracy parameter $a\in[0,1]$, the Prediction-Error-Robust Policy ${\pi^{\mathrm{PERP}}}$ achieves worst-case regret $$\mathcal{R}^{\pi^{\mathrm{PERP}}}(T)\leq \min\{3C_{\mathrm{\cref{upper bound on regret: past-demand-only}}} \sqrt{\log T}\cdot T^{(3+v)/4}, 2\ell \cdot T^a \},$$ where $C_{\mathrm{\cref{upper bound on regret: past-demand-only}}}$ is the constant in (ref) (and $2\ell$ matches the constant in (ref)a).

The intuition behind PERP is to follow the predictions until a time that is late enough to have evidence that the prediction quality is bad (compared to the Fixed-Time-Window Policy), while early enough to not incur much regret caused by the poor quality of the predictions. Because we do not observe the true past mean $\mu_t$ after time period $t$, we naturally use $\hat{\mu}_t^{\mathrm{fixed}}$ from the Fixed-Time-Window policy $\pi^{\mathrm{fixed}}$ as an estimation of $\mu_t$, and in turn keep tracking the cumulative difference the prediction quality $\left|a_t-\mu_t\right|$. We carefully choose the parameters in $\pi^{\mathrm{PERP}}$ so that this estimation is not accurate only with a small probability, and we can identify the prediction quality is bad if this cumulative difference is too large. By Proposition (ref), any policy can only achieve worst-case regret on the order of $T^{\min \{(3+v) / 4, a\}}$, so PERP is order-optimal.

\paragraph{ Aside: Unknown $v$ and Known $a$:} There are four possible scenarios depending on the knowledge of $v$ and $a$: known/unknown $v$ and known/unknown $a$. So far we have discussed three of them: known $v$ and $a$ ((ref)), unknown $v$ and $a$ ((ref)), and known $v$ and unknown $a$ ((ref)). For the sake of completeness, we discuss the remaining case of unknown $v$ and known $a$ in Appendix (ref), where we give a policy that achieves worst-case regret $\tilde{O}(T^{\min\{(3+v)/4, a\}})$. This is order-optimal by (ref).

Experiments on Real Data

Finally, we describe a set of experiments we performed to evaluate our policy (PERP) for the Nonstationary Newsvendor with Predictions. In all of our experiments, we compared PERP against the Shrinking-Time-Window Policy (NO-PRED) and the Prediction Policy (PURE-PRED). The main takeaways are:

enumerate• PERP's performance is robust with respect to the quality of the predictions, without knowing the prediction quality beforehand. Specifically, the (newsvendor) cost it incurs is consistently “close” to the lower of the costs incurred by NO-PRED and PURE-PRED. • PERP performs especially well when the absolute difference between the costs of NO-PRED and PURE-PRED is large, i.e. when the “stakes” are highest.

Experiments on Synthetic Data

The objective of our first batch of experiments was to fix one of the two theoretical parameters ($v$ or $a$) and test PERP's performance as the other parameters changes. To generate demand sequences, we used the parametric time-series model that corresponds to triple exponential smoothing (Holt Winters), a classic model for time-series data in the family of (ref). We give the exact formulas, and our choices of parameters, in Appendix (ref); for more on triple exponential smoothing, see winters1960forecasting. In our experiment, each demand sequence consisted of the demands for the next 365 time periods, with the realized demands generated as Poisson variables with the corresponding means. We ran two sets of experiments:

itemize• Fixed $v$: We fixed a single set of parameters for the demand sequence and generated 1,000 different predictions of this demand sequence, each from a set of “predicted” parameters with different accuracy. Thus the variation parameter $v$ was fixed, and the accuracy parameter $a$ varied across instances. • Fixed $a$: We generated 1,000 demand sequences by selecting the parameters randomly. We then generated predictions by changing each parameter 10% and using the corresponding sequence. Thus the variation parameter $v$ varied across instances, but the accuracy parameter $a$ was (roughly) fixed.
figure[figure omitted — 612 chars of source]

For each demand sequence and corresponding prediction, we ran NO-PRED, PURE-PRED, and PERP with equal overage and underage costs, and scaling constants $\kappa=\gamma=1$. Because the experiment was synthetic, the true underlying demand distribution was known at each time period. Therefore we also ran {\texttt OPT} as a benchmark, which simply ordered the optimal quantile at each time period. The variation parameter $v$ in PERP was calculated using the past demands of the pre-fixed 30 time periods by the definition in Section 2.2. We calculated the parameters $v$ and $a$ by their definitions (given in Section 2.2 and Section 4.1, respectively), scaled appropriately to make them lie in $[0,1]$. The resulted scatter plots are shown in (ref). In (a), $v$ is fixed, so the cost of NO-PRED (blue dots) is approximately the same for all instances. The cost of PURE-PRED (orange dots) is approximately exponential in $a$, which follows by (ref)a). In (b), $a$ is fixed, so the cost of \texttt{PURE-PRED} is approximately the same for all instances. The cost of \texttt{NO-PRED} is approximately exponential in $v$, which follows by (ref). Note that in both (a) and (b), the cost of \texttt{PERP} (green dots) is close to the minimal cost of \texttt{NO-PRED} and \texttt{PURE-PRED} across all instances, showing that \texttt{PERP}'s performance is robust in both $v$ and $a$.

Experiments on Real Data

We used real-world datasets to represent the “demand” sequences in our experiments. (ref) depicts example time series from each of these datasets. All datasets include {\em multiple} daily time series and are publicly available:

itemize• {\bf Rossmann:}\footnote{Available at \url{https://www.kaggle.com/competitions/rossmann-store-sales/data}} Daily number of customers that visited each of 1,115 stores in the {\em Rossmann} drug store chain during a 781-day period in 2013-2015. • {\bf Wikipedia:}\footnote{Available at \url{https://www.kaggle.com/competitions/web-traffic-time-series-forecasting/data}} Daily web traffic across {\em Wikipedia.com} pages of 9 different languages for an 803-day period from 2015 to 2017. • {\bf Restaurant:}\footnote{Available at \url{https://www.kaggle.com/competitions/recruit-restaurant-visitor-forecasting/}} Daily number of visitors and online reservations across 185 restaurants in Japan, during a 478-day period in 2016-2017. We treated the number of visitors as the “demand,” and the reservations as a predictive feature.
figure[figure omitted — 738 chars of source]

Each {\em instance} of our experiment represented a single Nonstationary Newsvendor with Predictions problem, with the realized demands taken from a single time series in our data (a single Rossmann store, a single language on Wikipedia, or a single restaurant). The overage and underage costs were constant within each instance, and without loss of generality the two costs for an instance can be characterized by the corresponding critical quantile (specifically the ratio of the underage cost to the sum). The time horizon for each instance was a set number of days taken from the end of the time series, with the preceding days used to train one of four prediction methods. These predictions were also updated over the course of the instance at a set frequency. For the Wikipedia dataset, this yielded a total of 2,880 possible instances, all of which were tested. The Rossmann dataset has multiple orders of magnitude more instances, so we randomly sampled 1,000 from this set. For the Restaurant dataset, we used a single prediction method to generate two sets of predictions for each restaurant: one only utilized the number of past visitors and the other incorporated the number of reservations as a feature, which gave 740 instances. (ref) describes all of the instances used.

table[table omitted — 609 chars of source]

For each instance, we applied NO-PRED, PURE-PRED, and PERP with scaling constants $\kappa=\gamma=1$, and the variation parameter $v$ in PERP was calculated using the past demands of the training data by the definition in Section 2.2. To generate predictions, we used four popular forecasting method ranging from classic to the state-of-the art:

itemize• {\bf Exponential Smoothing (Holt Winters):} A classic algorithm based on a (linear) trend and seasonality decomposition as in (ref), known for its simplicity and robust performance. It is frequently used as a benchmark in forecasting competitions (makridakis2000m3). Tuning parameters: seasonality of length 50. • {\bf ARIMA:} Another classic algorithm that is rich enough to model a wide class of nonstationary time-series. Tuning parameters: $(p,q,r)=(3,2,5)$. • {\bf Prophet:} A recent algorithm developed by Facebook (taylor2018forecasting) based on a (piecewise-linear) trend and seasonality decomposition as in (ref), known to work well in practice with minimal tuning. Tuning parameters: software default. • {\bf LightGBM:} A recent algorithm developed by Microsoft (ke2017lightgbm) based on tree algorithms. LightGBM formed the core of most of the top entries in the recent \$100,000 M5 Forecasting Challenge (makridakis2022m5). Tuning parameters: software default.

For the Restaurant dataset, we used Prophet as the forecasting method, with and without the reservations as an additive linear feature. We treated the outputs of these methods as predictions of the {\em mean} demand. To estimate the demand distribution around this mean, we used the empirical distribution of the residuals of the same predictions on the training period.\footnote{That is, if the training data consists of $T_{\mathrm{train}}$ periods, which without loss we index as $\{t=-T_{\mathrm{train}}+1,T_{\mathrm{train}}+2,\ldots,-1,0\}$, then the demand distribution at any time $t$ was estimated to be $$\hat{\mu}_t+\text{Uniform}\left(\{d_s - \hat{\mu}_s: s=-T_{\mathrm{train}}+1,T_{\mathrm{train}}+2,\ldots,-1,0\}\right).$$} In practice, even if the prediction quality is good, the predictions of the first few days might incur large costs due to noise/instability of the predictions, which may cause PERP to misidentify the prediction quality. Therefore we restricted PERP to following the predictions for the first 20 days, only allowing switches afterward.

\paragraph{\bf Results:} Each instance yields three total costs: one incurred by PERP, and two incurred by the benchmark algorithms (NO-PRED and PURE-PRED). The primary performance metric we report is a form of optimality gap. For an instance $I$, let $\mathrm{cost}^{\mathrm{PURE-PRED}}(I)$ be the cost of PURE-PRED, and similarly define $\mathrm{cost}^{\mathrm{NO-PRED}}(I)$, $\mathrm{cost}^{\mathrm{PERP}}(I)$. Then the optimality gap (GAP) of PERP is defined as $$\mathrm{GAP}(I)=\frac{\mathrm{cost}^{\mathrm{PERP}}(I)-\min\{\mathrm{cost}^{\mathrm{PURE-PRED}}(I),\mathrm{cost}^{\mathrm{NO-PRED}}(I)\}}{ \left|\mathrm{cost}^{\mathrm{PURE-PRED}}(I) - \mathrm{cost}^{\mathrm{NO-PRED}}(I)\right| }.$$ If we think of PERP as trying to achieve the minimum of the costs incurred by the two benchmark policies, then GAP measures the excess cost that \texttt{PERP} incurs on top of this minimum, normalized so that $\mathrm{GAP}=0$ implies that the minimum has been achieved, and $\mathrm{GAP}=1$ implies that the maximum of the two costs was incurred.\footnote{GAP may technically be outside of $[0,1]$.}

figure[figure omitted — 689 chars of source]

Experiments on the datasets yielded the histograms in (ref). For each instance $I$, the value on the horizonal axis is $\log(\mathrm{cost}^{\mathrm{PURE-PRED}}(I)/\mathrm{cost}^{\mathrm{NO-PRED}}(I))$, which is greater than 0 if NO-PRED has a lower cost, and less than 0 if PURE-PRED has a lower cost. In the 1,000 Rossman instances NO-PRED had a lower cost 82.7% of the time, in the 2,880 Wikipedia instances NO-PRED had a lower cost 81.9% of the time, and in the 740 Restaurant instances NO-PRED had a lower cost 64.3% of the time. The values on the vertical axis are the GAPs. Note that most GAPs are small when the absolute values of the log difference are large. This shows PERP performs very well when the difference of costs between \texttt{NO-PRED} and \texttt{PURE-PRED} is large. On the other hand, there are instances where \texttt{PERP} has large GAPs, in particular there are instances with GAPs equal to 1 when the log difference of costs is close to 0. This happens because when the log difference of costs is close to 0, the cost of \texttt{NO-PRED} and the cost of \texttt{PURE-PRED} are close, so \texttt{PERP} may misidentify the prediction quality. Still, since the max cost and the min cost of the other two policies are close, even the GAPs are large in these instances, \texttt{PERP} does not perform badly.

table[table omitted — 371 chars of source]

We further divide the instances according to which of NO-PRED and PURE-PRED had lower cost in (ref) and (ref).

figure[figure omitted — 1,134 chars of source]

For comparison, if we did not know the prediction quality beforehand, uniformly random choosing between NO-PRED and PURE-PRED has an expected GAP of 0.5. Therefore PERP outperforms this natural benchmark in all cases of all datasets.

Conclusion

We proposed a new model incorporating predictions into the nonstationary newsvendor problem. We first gave a complete analysis of the Nonstationary Newsvendor (without predictions) by proving a lower regret bound and developing the Shrinking-Time-Window Policy, which was the first policy that achieves the lower bound up to log factors without knowing the variation parameter. We then considered the Nonstationary Newsvendor with Predictions and proposed the Prediction-Error-Robust Policy, which does not need to know the prediction quality beforehand, and achieves nearly optimal minimax worst-cast regret.

thebibliography{99} \bibitem{an2024best} An, L., Li, A. A., Moseley, B., & Visotsky, G. (2024). Best of Many in Both Worlds: Online Resource Allocation with Predictions under Unknown Arrival Model. arXiv preprint arXiv:2402.13530. \bibitem{antoniadis2020secretary} Antoniadis, A., Gouleakis, T., Kleer, P., & Kolev, P. (2020). Secretary and online matching problems with machine learned advice. NeurIPS, 33, 7933-7944. \bibitem{arrow1958studies} Arrow, Kenneth Joseph, Samuel Karlin, Herbert E Scarf, and others, “Studies in the mathematical theory of inventory and production," 1958, Stanford University Press. \bibitem{auer2002nonstochastic} Auer, P., Cesa-Bianchi, N., Freund, Y., & Schapire, R. E. (2002). The nonstochastic multiarmed bandit problem. SIAM journal on computing, 32(1), 48-77. \bibitem{aviv2005partially} Aviv, Yossi and Amit Pazgal, “A partially observed Markov decision process for dynamic pricing," Management Science, 51(9), 1400--1416, 2005, INFORMS. \bibitem{azoury1985bayes} Azoury, Katy S, “Bayes solution to dynamic inventory models under unknown demand distribution," \textit{Management Science}, 31(9), 1150--1160, 1985, \textit{INFORMS}. \bibitem{balseiro2022single} Balseiro, Santiago, Christian Kroer, and Rachitesh Kumar, “Single-leg revenue management with advice," \textit{arXiv preprint arXiv:2202.10939}, 2022. \bibitem{ban2019big} Ban, Gah-Yi and Cynthia Rudin, “The big data newsvendor: Practical insights from machine learning," \textit{Operations Research}, 67(1), 90--108, 2019, \textit{INFORMS}. \bibitem{besbes2014stochastic} Besbes, Omar, Yonatan Gur, and Assaf Zeevi, “Stochastic multi-armed-bandit problem with non-stationary rewards," \textit{Advances in Neural Information Processing Systems}, 27, 2014. \bibitem{besbes2015non} Besbes, Omar, Yonatan Gur, and Assaf Zeevi, “Non-stationary stochastic optimization," \textit{Operations Research}, 63(5), 1227--1244, 2015, \textit{INFORMS}. \bibitem{besbes2013implications} Besbes, Omar and Alp Muharremoglu, “On implications of demand censoring in the newsvendor problem," \textit{Management Science}, 59(6), 1407--1424, 2013, \textit{INFORMS}. \bibitem{burnetas2000adaptive} Burnetas, Apostolos N. and Craig E. Smith, “Adaptive ordering and pricing for perishable products," \textit{Operations Research}, 48(3), 436--443, 2000, \textit{INFORMS}. \bibitem{chen2019coordinating} Chen, Boxiao, Xiuli Chao, and Hyun-Soo Ahn, “Coordinating pricing and inventory replenishment with nonparametric demand learning," \textit{Operations Research}, 67(4), 1035--1052, 2019, \textit{INFORMS}. \bibitem{cheung2022hedging} Cheung, W. C., Simchi-Levi, D., & Zhu, R. (2022). Hedging the drift: Learning to optimize under nonstationarity. \textit{Man. Sci.}, 68(3), 1696-1713. \bibitem{dutting2021secretaries} Dütting, Paul, Silvio Lattanzi, Renato Paes Leme, and Sergei Vassilvitskii, “Secretaries with advice," in \textit{Proceedings of the 22nd ACM Conference on Economics and Computation}, pp. 409--429, 2021. \bibitem{edgeworth1888mathematical} Edgeworth, Francis Y., “The mathematical theory of banking," \textit{Journal of the Royal Statistical Society}, 51(1), 113--127, 1888. \bibitem{godfrey2001adaptive} Godfrey, Gregory A. and Warren B. Powell, “An adaptive, distribution-free algorithm for the newsvendor problem with censored demands, with applications to inventory and distribution," \textit{Management Science}, 47(8), 1101--1112, 2001, \textit{INFORMS}. \bibitem{hao2023leveraging} Hao, B., et al. (2023). Leveraging demonstrations to improve online learning: Quality matters. In \textit{ICML}. \bibitem{hu2024constrained} Hu, P., Jiang, J., Lyu, G., & Su, H. (2024). Constrained online two-stage stochastic optimization: Algorithm with (and without) predictions. \textit{arXiv preprint arXiv:2401.01077}. \bibitem{huber2019data} Huber, Jakob, Sebastian Müller, Moritz Fleischmann, and Heiner Stuckenschmidt, “A data-driven newsvendor problem: From data to decision," \textit{European Journal of Operational Research}, 278(3), 904--915, 2019, \textit{Elsevier}. \bibitem{huh2009nonparametric} Huh, Woonghee Tim and Paat Rusmevichientong, “A nonparametric asymptotic analysis of inventory planning with censored demand," \textit{Mathematics of Operations Research}, 34(1), 103--123, 2009, \textit{INFORMS}. \bibitem{iglehart1964dynamic} Iglehart, Donald L, “The dynamic inventory problem with unknown demand distribution," \textit{Management Science}, 10(3), 429--440, 1964, \textit{INFORMS}. \bibitem{jin2022online} Jin, B., & Ma, W. (2022). Online bipartite matching with advice: Tight robustness-consistency tradeoffs for the two-stage model. \textit{Advances in Neural Information Processing Systems}, 35, 14555-14567. \bibitem{karlin1960dynamic} Karlin, Samuel, “Dynamic inventory policy with varying stochastic demands," \textit{Management Science}, 6(3), 231--258, 1960, \textit{INFORMS}. \bibitem{karnin2016multi} Karnin, Z. S., & Anava, O. (2016). Multi-armed bandits: Competing with optimal sequences. \textit{NIPS}, 29. \bibitem{ke2017lightgbm} Ke, Guolin, Qi Meng, Thomas Finley, Taifeng Wang, Wei Chen, Weidong Ma, Qiwei Ye, and Tie-Yan Liu, “Lightgbm: A highly efficient gradient boosting decision tree," \textit{Advances in Neural Information Processing Systems}, 30, 2017. \bibitem{keskin2021nonstationary} Keskin, N. Bora, Xu Min, and Jing-Sheng Jeannette Song, “The nonstationary newsvendor: Data-driven nonparametric learning," \textit{Available at SSRN 3866171}, 2023. \bibitem{keskin2017chasing} Keskin, N. Bora and Assaf Zeevi, “Chasing demand: Learning and earning in a changing environment," \textit{Mathematics of Operations Research}, 42(2), 277--307, 2017, \textit{INFORMS}. \bibitem{kleywegt2002sample} Kleywegt, Anton J., Alexander Shapiro, and Tito Homem-de-Mello, “The sample average approximation method for stochastic discrete optimization," \textit{SIAM Journal on Optimization}, 12(2), 479--502, 2002, \textit{SIAM}. \bibitem{kunnumkal2008using} Kunnumkal, Sumit and Huseyin Topaloglu, “Using stochastic approximation methods to compute optimal base-stock levels in inventory control problems," \textit{Operations Research}, 56(3), 646--664, 2008, \textit{INFORMS}. \bibitem{LattanziLMV20} Lattanzi, Silvio, Thomas Lavastida, Benjamin Moseley, and Sergei Vassilvitskii, “Online Scheduling via Learned Weights," in \textit{Proceedings of the 2020 ACM-SIAM Symposium on Discrete Algorithms (SODA 2020)}, Salt Lake City, UT, USA, January 5-8, 2020, 1859--1877, \textit{SIAM}. \bibitem{lattanzi2020online} Lattanzi, Silvio, Thomas Lavastida, Benjamin Moseley, and Sergei Vassilvitskii, “Online scheduling via learned weights," in \textit{Proceedings of the Fourteenth Annual ACM-SIAM Symposium on Discrete Algorithms}, 1859--1877, 2020, \textit{SIAM}. \bibitem{levi2015data} Levi, Retsef, Georgia Perakis, and Joline Uichanco, “The data-driven newsvendor problem: New bounds and insights," \textit{Operations Research}, 63(6), 1294--1306, 2015. \bibitem{levi2007provably} Levi, Retsef, Robin O. Roundy, and David B. Shmoys, “Provably near-optimal sampling-based policies for stochastic inventory control models," \textit{Mathematics of Operations Research}, 32(4), 821--839, 2007. \bibitem{liyanage2005practical} Liyanage, Liwan H and J. George Shanthikumar, “A practical inventory control policy using operational statistics," \textit{Operations Research Letters}, 33(4), 341--348, 2005, \textit{Elsevier}. \bibitem{lovejoy1990myopic} Lovejoy, William S, “Myopic policies for some inventory models with uncertain demand distributions," \textit{Management Science}, 36(6), 724--738, 1990, \textit{INFORMS}. \bibitem{luo2018efficient} Luo, H., Wei, C. Y., Agarwal, A., & Langford, J. (2018). Efficient contextual bandits in non-stationary worlds. In \textit{Conference On Learning Theory} (pp. 1739-1776). PMLR. \bibitem{DBLP:journals/jacm/LykourisV21} Lykouris, Thodoris and Sergei Vassilvitskii, “Competitive Caching with Machine Learned Advice," \textit{Journal of the ACM (JACM)}, 68(4), 24:1--24:25, 2021, \textit{ACM New York, NY}. \bibitem{lykouris2021competitive} Lykouris, Thodoris and Sergei Vassilvitskii, “Competitive caching with machine learned advice," \textit{Journal of the ACM (JACM)}, 68(4), 1--25, 2021, \textit{ACM New York, NY}. \bibitem{mahdian2012online} Mahdian, M., Nazerzadeh, H., & Saberi, A. (2012). Online optimization with uncertain information. \textit{ACM Transactions on Algorithms (TALG)}, 8(1), 1-29. \bibitem{makridakis2000m3} Makridakis, Spyros and Michele Hibon, “The M3-Competition: results, conclusions, and implications," \textit{International Journal of Forecasting}, 16(4), 451--476, 2000. \bibitem{makridakis2022m5} Makridakis, Spyros, Evangelos Spiliotis, and Vassilios Assimakopoulos, “M5 accuracy competition: Results, findings, and conclusions," \textit{International Journal of Forecasting}, 38(4), 1346--1364, 2022. \bibitem{MitzenmacherV22} Mitzenmacher, Michael and Sergei Vassilvitskii, “Algorithms with Predictions," \textit{Communications of the ACM (CACM)}, 65(7), 33--35, 2022. \bibitem{munoz2017revenue} Munoz, Andres and Sergei Vassilvitskii, “Revenue optimization with approximate bid predictions," \textit{Advances in Neural Information Processing Systems}, 30, 2017. \bibitem{oroojlooyjadid2020applying} Oroojlooyjadid, Afshin, Lawrence V Snyder, and Martin Takáč, “Applying deep learning to the newsvendor problem," \textit{IISE Transactions}, 52(4), 444--463, 2020, \textit{Taylor & Francis}. \bibitem{powell2004learning} Powell, Warren, Andrzej Ruszczyński, and Huseyin Topaloglu, “Learning algorithms for separable approximations of discrete stochastic optimization problems," \textit{Mathematics of Operations Research}, 29(4), 814--836, 2004, \textit{INFORMS}. \bibitem{rohatgi2020near} Rohatgi, Dhruv, “Near-optimal bounds for online caching with machine learned advice," in \textit{Proceedings of the Fourteenth Annual ACM-SIAM Symposium on Discrete Algorithms}, 1834--1845, 2020, \textit{SIAM}. \bibitem{scarf1959bayes} Scarf, Herbert, “Bayes solutions of the statistical inventory problem," \textit{The Annals of Mathematical Statistics}, 30(2), 490--508, 1959, \textit{JSTOR}. \bibitem{scarf1960optimality} Scarf, Herbert, K. Arrow, S. Karlin, and P. Suppes, “The optimality of (S, s) policies in the dynamic inventory problem," in \textit{Optimal pricing, inflation, and the cost of price adjustment}, pp. 49--56, 1960, \textit{MIT Press Cambridge}. \bibitem{shapiro2003monte} Shapiro, Alexander, “Monte Carlo sampling methods," \textit{Handbooks in Operations Research and Management Science}, 10, 353--425, 2003, \textit{Elsevier}. \bibitem{taylor2018forecasting} Taylor, Sean J. and Benjamin Letham, “Forecasting at scale," \textit{The American Statistician}, 72(1), 37--45, 2018. \bibitem{treharne2002adaptive} Treharne, James T. and Charles R. Sox, “Adaptive inventory control for nonstationary demand and partial information," \textit{Management Science}, 48(5), 607--624, 2002, \textit{INFORMS}. \bibitem{vershynin2018high} Vershynin, Roman, “High-dimensional probability: An introduction with applications in data science," \textit{Cambridge University Press}, vol. 47, 2018. \bibitem{winters1960forecasting} Winters, P. R. (1960). Forecasting sales by exponentially weighted moving averages. \textit{The Use of MMR, Diversity-Based Reranking for Reordering Documents and Producing Summaries}, 6(3), 324-342. \bibitem{zhang2023optimal} Zhang, Luhao, Jincheng Yang, and Rui Gao, “Optimal robust policy for feature-based newsvendor," \textit{Management Science, Forthcoming}, 2023. \bibitem{baby2019online} Baby D, Wang YX (2019) Online forecasting of total-variation-bounded sequences. \emph{Advances in Neural Information Processing Systems} 32. \bibitem{chen2019nonstationary} Chen X, Wang Y, Wang YX (2019) Nonstationary stochastic optimization under L p, q-variation measures. \emph{Operations Research} 67(6):1752--1765. \bibitem{huang2023stability} Huang C, Wang K (2023) A stability principle for learning under non-stationarity. \emph{arXiv preprint arXiv:2310.18304}. \bibitem{yang2016tracking} Yang T, Zhang L, Jin R, Yi J (2016) Tracking slowly moving clairvoyant: Optimal dynamic regret of online learning with true and noisy gradient. \emph{International Conference on Machine Learning}, 449--457 (PMLR). \bibitem{zhang2018adaptive} Zhang L, Lu S, Zhou ZH (2018) Adaptive online learning in dynamic environments. \emph{Advances in Neural Information Processing Systems} 31. \bibitem{bai2022adapting} Bai Y, Zhang YJ, Zhao P, Sugiyama M, Zhou ZH (2022) Adapting to online label shift with provable guarantees. \emph{Advances in Neural Information Processing Systems} 35:29960--29974.