Advertising is considered as an important strategic tool to promote product and improve sales. Extensive research has been devoted to advertising strategies and their effect on product sales. Chiefly because aggregate-level sales data are easy to collect, the prior studies predominately develop and analyze aggregate advertising models which relate product sales to advertising spending under a known sales response function. Nowadays, however, the emergence of Internet, e-commerce and data analytics approaches has made collecting data on individual consumer behavior and real-time sales feasible. Therefore, studying more sophisticated advertising models which can exploit these data is necessary and meaningful. In this dissertation, we consider two dynamic advertising models, one incorporates customer satisfaction and customer switching behavior and the other involves dynamic sales learning. / The first model focuses on the markets of experience goods whose quality levels are unobservable to the buyers. The buyers make the purchase decisions based on their past usage experience of the goods and the advertising outlays of the sellers. We first consider the competitive market where there are multiple brands planning their advertising campaigns. We derive the long-term steady-state equilibrium advertising strategies and market shares of the brands. We study how customer reaction to their past usage experience of the product (satisfaction) affects the sellers’ advertising strategies and market shares. We further analyze the monopoly market, where the focus is on the question of whether the monopolist should use even-level advertising or pulsing advertising strategy. / In the second model, we study the dynamic advertising budget allocation problems, in which the relationship between the advertising expenditure and the product sales is unknown to the retailer and the retailer can only learn this information through observing realized sales. We propose nonparametric advertising budget allocation policies for both single- and multi-product problems. We show that such policies are asymptotically optimal. In particular, for the single-product problem, by constructing a lower-bound instance, we show that our policy achieves near-best asymptotic performance. / 广告预算的确定和分配是企业运营中一个极为重要的决策。而广告资金的动态支出策略已经被研究了数十年。由于以前可获取的数据往往仅限于市场的一些宏观数据。传统文献主要用一个已知的销售响应函数从宏观层面刻画广告支出和销售量之间的关系。现今，随着互联网，电子商务，社交媒体和数据分析方法的出现，使收集有关消费者行为和实时销售量的数据成为可能。合理地利用这些数据可以大大提高企业的广告效率。这也给广告策略的研究带来了新的契机。本论文研究两个动态广告支出策略模型，一个模型涉及顾客满意度和顾客转换行为，另外一个涉及销售响应函数的动态学习。 / 本文的第一个模型考虑体验商品市场中卖家的广告支出策略。在市场中，顾客没有办法观测到产品的真实质量，他们的产品选择受到自己之前的产品体验和卖家的广告支出的影响。我们首先考虑有竞争的市场，市场中有多个品牌同时进行广告支出决策。我们推导了市场的长期稳态平衡。研究了不同的客户满意度与顾客转换行为的关系对卖家的广告支出策略和市场份额的影响。然后，我们分析了垄断市场中垄断卖家的长期广告支出策略。此外，我们还讨论了该卖家应该使用持续的广告投入策略还是周期性脉冲广告策略的问题。 / 本文的第二个模型研究是的动态广告预算分配问题。我们假设卖家起初并不知道广告支出的销售响应函数，他只能通过观察实时销售数据来对销售响应函数进行学习。我们分别考虑了单产品和多产品的问题，并提出了相应的非参数动态广告预算分配策略。我们证明了所提出的动态预算分配策略是渐进最优的。对于单产品的问题，通过构造出一类“最坏”的响应函数，我们证明了所提出的动态预算分配策略的渐进绩效已基本接近最优。 / Yang, Chaolin. / Thesis Ph.D. Chinese University of Hong Kong 2015. / Includes bibliographical references (leaves 133-140). / Abstracts also in Chinese. / Title from PDF title page (viewed on 12, October, 2016). / Detailed summary in vernacular field only. / Detailed summary in vernacular field only. / Detailed summary in vernacular field only.

Advertising is the economic engine of the internet. It allows online platforms to fund services that are free at the point of use, while providing businesses the opportunity to target their ads at relevant users. The mechanism of choice for selling these advertising opportunities is real-time auctions: whenever a user visits the platform, an auction is run among interested advertisers, and the winner gets to display their ad to the user. The entire process runs in milliseconds and is implemented via automated algorithms which bid on behalf of the advertisers in every auction. These automated bidders take as input the high-level objectives of the advertiser like value-per-click and budget, and then participate in the auctions with the goal of maximizing the utility of the advertiser subject to budget constraints. Thus motivated, this thesis develops a theory of bidding in auctions under budget constraints, with the goal of informing the design of automated bidding algorithms and analyzing the market-level outcomes that emerge from their simultaneous use. First, we take the perspective of an individual advertiser and tackle algorithm-design questions. How should one bid in repeated second-price auctions subject to a global budget constraint? What is the optimal way to incorporate data into bidding decisions? Can data be incorporated in a way that is robust to common forms of variability in the market? As we analyze these questions, we go beyond the problem of bidding under budget constraints and develop algorithms for more general online resource allocation problems. In Chapter 2, we study a non-stationary stochastic model of sequential auctions, which despite immense practical importance has received little attention, and propose a natural algorithm for it. With access to just one historical sample per auction/distribution, we show that our algorithm attains (nearly) the same performance as that possible under full knowledge of the distributions, while also being robust to distribution shifts which typically occur between the sampling and true distributions. Chapter 3 investigates the impact of uncertainty about the total number of auctions on the performance of bidding algorithms. We prove upper bounds on the best-possible performance that can be achieved in the face of such uncertainty, and propose an algorithm that (nearly) achieves this optimal performance guarantee. We also provide a fast method for incorporating predictions about the total number of auctions into our algorithm. All of our proposed algorithms implement some version of FTRL/Mirror-Descent in the dual space, making them ideal for large-scale low-latency markets like online advertising. Next, we look at the market as a whole and analyze the equilibria which emerge from the simultaneous use of automated bidding algorithms. For example, we address questions like: Does an equilibrium always exist? How does the auction format (first-price vs second-price) impact the structure of the equilibria? Do automated bidding algorithms always efficiently converge to some equilibrium? What are the social welfare properties of these equilibrium outcomes? We systematically examine such questions using a variety of tools, ranging from infinite-dimensional fixed-point arguments for proving existence of structured equilibria, to computational complexity results about finding them. In Chapter 4, we start by establishing the existence of equilibria based on pacing—a practically-popular and theoretically-optimal budget management strategy—for all standard auctions, including first-price and second-price auctions. We then leverage its structure to establish a revenue equivalence result and bound the price of anarchy of liquid welfare. Chapter 5 looks at the market from a computational lens and investigates the complexity of finding pacing-based equilibria. We show that the problem is PPAD complete, which in turn implies the impossibility of polynomial-time convergence of any pacing-based automated bidding algorithms (under standard complexity-theoretic assumptions). Finally, in Chapter 6, we move beyond pacing-based strategies and investigate throttling, which is another popular method for managing budgets in practice. Here, we describe a simple tâtonnement-style algorithm which efficiently converges to an equilibrium in first-price auctions, and show that no such algorithm exists for second-price auctions (under standard complexity-theoretic assumptions). Furthermore, we prove tight bounds on the price of anarchy for liquid welfare, and compare platform revenue under throttling and pacing.