%Paper: ewp-io/9309001 %From: Neil Gandal %Date: Mon, 13 Sep 93 13:57:19 IST \documentstyle[12pt]{article} \textwidth=437.27pt \topmargin=0in \textheight=8in \oddsidemargin=23pt \reversemarginpar %\renewcommand\baselinestretch{1.5} %the above command stretches all the paper including the footnotes \renewcommand{\thesection}{\arabic{section}.} \renewcommand{\thesubsection}{\thesection{\arabic{subsection}}} \let\Large=\large \let\large=\normalsize \pagestyle{myheading}\markright{NEVER WORKS} \begin{document} \baselineskip=0.3in \newtheorem{defn}{DEFINITION} \newtheorem{ass}{Assumption} \newtheorem{lem}{Lemma} \newtheorem{theo}{Theorem} \newtheorem{cor}{Corollary} \newtheorem{prop}{Proposition} \def\l{{\lambda}} \def\t{{\tau}} \def\p{{\pi}} \def\g{{\gamma}} \begin{titlepage} \vspace*{0.3in} \begin{center} {\Large\bf Experimentation and Learning with Network Effects} \\ \vspace{.4in} Arthur Fishman \\ Tel Aviv University \\ and \\ Neil Gandal\\ Tel Aviv University \\ \vspace{0.2in}\today \\ \vspace{.2in}{\bf Abstract}\\ \end{center} %\marginpar{draft=dur17.tex} This paper considers learning in an imperfectly competitive setting. By allowing an opponent a ``head start," unsuccessful unilateral experimentation may jeopardize future sales and profits. We show that even in the absence of spillover and signalling effects, competition can inhibit the scope of learning, relative to a monopoly. \vspace{4 mm} \noindent {\bf JEL Classification Numbers:} D83 \vspace{0.1in}\noindent \end{titlepage} \baselineskip=0.2in \section{Introduction} \hspace*{6 mm} Consider a seller who is imperfectly informed about market demand. She can improve her information by experimenting with prices. However, experimentation is costly: charging a high price generates valuable information, but will reduce current sales and profits if demand is actually low. Therefore the seller's optimal learning policy must balance the costs of experimentation against the potential gains. In a seminal paper, Rothschild (1974) examined the optimal dynamic learning strategy of a monopoly seller. He showed that the cost of experimentation will typically inhibit complete learning, even though through sufficiently extensive experimentation, the true state of demand can be learned with arbitrarily high probability. More recent studies of the optimal learning problem include McLennan (1984), Easley and Kiefer (1988), Creane (1989), Mirman, Samuelson, and Urbano (1989) and Aghion, Bolton, Harris and Jullien (1991). All of these papers considered a monopoly seller, i.e., a single optimizing decision maker. In this paper, we consider equilibrium learning in an imperfectly competitive setting. In our setting, learning involves in addition to the costs considered in the existing literature, a novel cost. By allowing an opponent a ``head start," unsuccessful unilateral experimentation involves not only the conventional loss of current profits, but may jeopardize future sales and profits. In settings where initial market shares are important for generating future profits, unilateral experimentation at the market's inception may lead to the loss of current and {\bf future} sales. Settings in which this effect may be important include markets characterized by network externalities, switching costs or lock-in, and learning by doing. In a competitive setting, there may be additional factors impinging upon the decision to experiment. For example, Rob (1991) has shown that spillover effects may reduce the amount of learning. Similarly signalling effects might also distort the learning process. For example, a firm might charge low prices to convince a potential entrant that demand is low, further inhibiting learning. Aghion, Espinosa, and Jullien (1990) and Mirman, Samuelson, and Schlee (1991) explore the incentives for strategic manipulation of information in duopoly settings. In this paper, we present a simple model with two decision makers that allows us to isolate a ``market share" effect, i.e., the framework has been constructed so that spillover and signalling effects are absent. In this setting, we show that there is a collectively suboptimal amount of learning which is a direct consequence of competition. In other words, there exists a set of parameters for which a monopoly seller would experiment and learn about demand, but in the competitive setting, the unique subgame perfect equilibrium is for neither firm to experiment: We term this lack of lack of learning ``the strategic Rothschild effect." The message of the paper is thus that competition may exacerbate the Rothschild effect.\footnote{Smith (1991) finds that the Rothschild effect is mitigated by the presence of multiple decision makers. His result is due to the absence of any strategic interaction among agents: In his setting the payoff of each agent is independent of the actions of other agents. Our goal is precisely to address the inhibitory strategic effects on equilibrium learning associated with multiple {\bf competing} agents.} \section{An Experimentation Model} \hspace*{6 mm} The market opens in two periods. In each period, a new consumer cohort enters the market. The size of each cohort is normalized to one. Consumers are homogeneous and demand one unit, at the most. Consumers knows their reservation values, but the firms are ex ante uncertain whether demand is high or low: each firm assigns the prior $q$ ($1-q)$ that the common consumer reservation value is $v_h$ ($v_l)$, where $v_h>v_l$. There are two firms, each of whom can supply an unlimited amount of the product at zero cost. The firms are risk neutral, and there is no discounting. Consumers must make actual contact with a seller in order to learn its price and it is assumed to be too costly to contact two firms in the same period. This grants the firms market power to set prices.\footnote{Actually, in the framework considered here, even a small search cost will be sufficient to grant firms full market power with respect to pricing. See Diamond (1971).} Therefore first period consumers choose a firm at random, and buy from that firm if the price does not exceed the reservation value.\footnote{Consumers are not dynamically strategic, i.e., they do not delay purchasing in the hope of realizing a lower future price. This is a common assumption in the network literature. See Katz and Shapiro (1986) for example.} Thus, in the presence of complete information about demand, firms would have full monopoly power. As discussed in the following paragraph, the only competitive ``liability" arises from the strategic inhibition to learn about demand, i.e., to experiment in prices in the markets' formative stage. In the second period, there is a network effect based on the firms' realized first period market shares. We make the simplifying assumption that the network effect is so important that no second period consumer is willing to purchase from the firm with a smaller installed base at any positive price.\footnote{Our results remain qualitatively unchanged as long as there is some network effect.} Thus if a firm realizes a smaller market share in the first period, it faces zero demand in the second period. If the firms achieve equal first period market shares, second period demand for each firm is identical to that of the first period. At the beginning of the second period, firms simultaneously set prices. We assume that they do so before learning which firm has achieved the larger first period market share or learning their rival's first period price. This assumption eliminates the presence of any learning externalities, i.e., it eliminates the possibility of free-riding on learning achieved by one's rival. Thus the case for experimentation is as compelling as possible. The firms are initially imperfectly informed about market demand. Clearly this information is of value to them. Specifically, under conditions of complete information about demand, each firm would charge the true reservation price in each period. This is because, by assumption, each first period consumer accepts any price not exceeding its reservation price. Thus by charging the actual reservation price, a firm is assured of not suffering any adverse networks effects in the future. Conversely, it cannot obtain any network advantage by charging less than this price. Therefore, in the absence of any incomplete information about demand, each firm would enjoy complete monopoly power over its segment of the market. Given incomplete information, a firm can become perfectly informed by charging the high price in the first period: acceptance of price $v_h$ perfectly reveals that demand is high, while its rejection reveals that demand is low. This represents the potential benefits from experimentation. Experimentation is costly, however. In our framework, this cost has two components. The first is the standard one, the potential loss of sales in the first period. The second and novel component, which our model is constructed to address in the starkest possible way, is a {\bf strategic} cost. The source of this cost is the loss of future sales via an adverse network effect, which occurs if unilateral experimentation causes that firm to realize a smaller first period market share. A natural benchmark which is useful to assess the significance of the second component, is comparison with monopoly learning in the same context. A monopoly also faces the first cost associated with experimentation. However, the strategic cost is absent because there is no rival who can capture its future sales. Because there is no scope for any {\bf direct} price competition in our framework, the absence of a network effect would effectively transform each firm into a monopoly. In the following section, we derive conditions under which learning is the monopolist's optimal strategy. This serves a benchmark with respect to which we subsequently identify the effect of competition on the optimal learning strategy. \section{Optimal Monopoly Learning} \hspace*{6 mm} If the monopolist chooses to experiment in the first period, i.e., if it charges $v_h$, it becomes perfectly informed. If the consumer does not purchase in the first period, the monopoly concludes that the reservation price is $v_l$ and charges this price in the second period. Otherwise $v_h$ is charged in the second period. From the perspective of the first period, the expected second period profits are $qv_h+(1-q)v_l$. Thus the expected total profits from experimenting (charging price $v_h$ in the first period) are $qv_h + (qv_h+(1-q)v_l)$, where $qv_h$ are the expected first period profits. Therefore a monopolist will experiment if and only if $qv_h + (qv_h+(1-q)v_l) \geq 2v_l$, where $2v_l$ are the total profits earned by the monopolist if it does not experiment in the first period and hence charges $v_l$ in both periods.\footnote{If the monopolist has no incentive to experiment in the first period, it will not experiment in the second period either.} Rearranging terms in the above equation, the monopolist will experiment if and only if \begin{equation} q \geq {v_l \over (2v_h-v_l)} \equiv q^m. \label{mon2} \end{equation} %\marginpar{mon2} If $q2qv_h$, which is the case for the q under consideration. We will now show that $v_l$ is also a best response to $v_h$. From the analysis of case I, if firm $i$ charges $v_l$ in the first period, it will also charge $v_l$ in the second period if $q< {2v_l \over (v_h+v_l)}$, which is implied by the assumption that $q