آموزش بهینه و حباب های فن آوری های نوین

It is widely believed that there is a fundamental linkage between major technological innovations, speculative fever, and wasteful overinvestment. This paper presents an equilibrium model of investment in a new industry, whose return-to-scale is not known in advance. Overinvestment relative to the full-information case is then optimal as the most efficient way to learn about the new technology. Moreover, the initial overinvestment is accompanied by apparently inflated stock prices and apparently negative expected excess returns in the new industry, which are also fully rational. This suggests a new interpretation of what seems to be stock market driven real bubbles.

It has recently become widely believed that technological breakthroughs inevitably entail economic excess. The pattern of the recent IT-driven boom and bust has led many commentators and historians to note parallels with earlier technological revolutions—railroads, canals, electric power—which ignited a burst of apparent overbuilding by, and apparent overvaluation of, innovating firms.1 These rapid expansions were all followed by longer adjustment phases during which the initial excesses were damped toward long-run equilibrium values. The description does not fit all waves of innovation or all financial bubbles. But the recurrence of the pattern does raise the question of what it is about the technological innovations that induced these dynamics. From a business cycle perspective, such episodes are unusual for a number of reasons. In general, overinvestment in response to a technology shock, large or small, is not a feature of standard models. Likewise, the subsequent disinvestment—without any technological regress—is hard to explain as an optimizing policy. Moreover, these responses reverse the normal asymmetry in which sharp recessions are followed by gradual expansions. Similarly reversed is the usual pattern of investment seeming to respond too strongly to cash-flow and not strongly enough to Tobin's q. From an asset pricing perspective, any pattern of apparently predictable negative returns is also very difficult to explain. This paper suggests one mechanism that can account for these facts. It studies the short-run equilibrium dynamics following the introduction of a new production technology in a standard equilibrium setting. In this context, I show that, when the return-to-scale of the new technology are not known a priori, optimal policies can feature both initial overshooting of real investment and predictable deflation in the price of claims to the new sector. This behavior is driven by the incentive to efficiently learn the curvature of the production function—and hence the optimal long-run scale of the new industry—about which agents are uncertain. Indeed, this particular type of uncertainty could be said to be the distinguishing feature of a truly revolutionary technology: there is no historical experience of how it will scale up. No one knows how the interplay of competition, regulation, and costs will work out at vastly greater levels of production than have ever been seen before. In general, adaptive learning models can induce either caution or experimentation. In my formulation, agents have an incentive to push investment beyond the level that would seem optimal with full information in order to efficiently learn the shape of the production function. As experience grows, this incentive diminishes and investment declines. Market prices of installed capital mirror the gains to be had from learning. Tobin's q for the new industry starts high and then predictably subsides. The model is not intended as a general theory of booms and busts. Nor does it attempt to model either the evolution of the new technology or the process of its adoption. 2 Instead, the goal is to focus on the apparent overshooting of both real and financial quantities that seems to have characterized several important historical periods. Given the enormous literature on financial bubbles and the still larger one on technology-driven business cycles, it is not surprising that alternative explanations for real bubbles already exist. The simplest explanation is just error. Growth rate expectations drive levels of prices and investment, and expectations can be wrong. For important innovations, growth rates themselves are large numbers. Hence small errors can have big consequences. Here two distinct perspectives can be taken. On one hand, the episodes that stand out over the course of history may simply be the most visible instances of what is essentially idiosyncratic error. In this view, overestimated growth rates lead to spectacular bubbles, while underestimated ones, equally often, lead to unremarkable gradual adjustments. On the other hand, the errors could be systematic. This stance has been forcefully argued by Shiller (2000), who links historical “new economy” sentiment to persistent cognitive biases generating irrational exuberance. The behavioral finance literature has built an impressive body of evidence documenting biases in expectations and systematic negative returns to high-growth, new, and low book-to-market stocks. Some recent empirical work (Polk and Sapienza, 2002 and Gilchrist et al., 2002) supports the notion of a behavioral link to real investment. It is worth noting that the behavioral argument does not per se require a role for the stock market. Irrationally exuberant agents would presumably overinvest however allocations were implemented. That said, undoubtedly the most widely held view of the recent boom/bust does involve a melding of irrationality with an accelerator-type role for financial markets. The following quote, from the IMF's 2001 World Economic Outlook and written by distinguished economists, certainly reflects a broad public perception. As in past technological revolutions, the initial phase of the IT revolution appears to have been characterized by excessive optimism about the potential earnings of innovating firms. This over-optimism led for several years to soaring stock prices of IT firms, which made equity finance cheaper and more readily available, which in turn boosted investment by IT firms. Equilibrium models in which financial markets amplify fundamental shocks represent another class of potential explanation for real bubbles, not necessarily related to irrationality. Models such as Carlstrom and Fuerst (1997), and Bernanke et al. (1999) instead embed contracting frictions which lead to endogenous variation in the cost of external finance. A large related literature explores the implications of frictions in credit markets (Holmstron and Tirole, 1997 and Allen and Gale, 2000). These models can explain excessive real disinvestment in recessions, when financial constraints bind. Contracting problems can also directly influence investment without reliance on a financial mechanism. In Philippon (2003) real business cycles are amplified by endogenous loosening of corporate governance in expansions. Models of social learning can produce herding by managers when private information is noncontractible. Scharfstein and Stein (1990) and Caplin and Leahy (1994) have used this approach to model overinvestment. In a similar vein, what might be called local learning models study the properties of economies populated by agents who adapt their behavior in sensible, but not formally optimal, fashion given their experience. This too can lead to overinvestment cascades.3 These models bridge the gap between rational and behavioral perspectives. As this brief summary makes clear, there are already a variety of devices that can be used to link technological advances to overinvestment. Relative to these, the model presented here is notably different in not relying on any form of market failure or frictions. The point of this distinction is not to be doctrinaire: undoubtedly failures and frictions play important roles in the type of episodes under consideration (and the model here does not preclude these). Instead, the goal is to point out that the conventional understanding of new technology bubbles may be incomplete. Moreover, if irrational exuberance and financial amplification are not the whole story, then the common interpretation of these episodes as wastefully misallocating resources may be misguided. In the context of this paper's model, in fact, boom followed by bust is the quickest (and most efficient) adjustment path to the long-run optimum. In modeling learning as an active process to be optimally managed, this paper follows a line of literature begun by Prescott (1972), who first considered the problem of stochastic control when the control affects the information set.4 No separation principle applies in these settings: that is, the problem cannot be decomposed into separate estimation and optimization stages. This makes analytical solutions impossible except in highly simplified settings. Moreover, even in one- and two-period problems the effects of learning incentives can be complex and ambiguous. In some cases, the intuition that learning can motivate active experimentation (e.g. via increased output or investment) is validated. In others, the opposite intuition holds: the need to learn can induce caution and waiting while knowledge accumulates.5 A two-period growth model similar to mine is analyzed by Bertocchi and Spagat (1998) who observe regions of both underinvestment and overinvestment relative to the full-information case. The scope for drawing policy conclusions for actual economic problems has thus been quite limited.6 More recently, increases in computing power have enabled analysis of the effects of optimal learning in more realistic settings. Wieland (2000b), for example, is able to deduce important implications for a monetary authority learning about an unknown money demand function while also controlling inflation. My aims are similar in scope: to be able to address the quantitative consequences of adaptive control in a dynamic equilibrium, incorporating standard utility functions, non-trivial constraints, and multiple periods. The model is still stylized and incomplete, yet it is able to offer some significant insights into how and when learning can induce bubble-like dynamics in the real economy. The paper is organized as follows. The next section gives the details of the economic setting. The information structure and the optimization problem are described, and the weaknesses and driving assumptions of the model are discussed. Section 3 presents solutions which establish the occurrence of the overinvestment effect and demonstrate that asset prices as well as investment become inflated. Section 4 examines patterns of returns that can arise under the model, and relates these to the empirical finance literature. The effects produced in the model can be large enough to account for the anomalous returns observed in new or high growth stocks. On this basis, the model appears consistent with available measurement of price bubbles. The final section briefly summarizes the paper's contribution.

This study has been motivated by the intriguing parallels between the recent IT boom/bust and earlier historical technological revolutions. There appears to be widespread public acceptance of two “stylized facts” about these parallels: First, that there is something inevitable about the linkage between innovation and bubbles (exemplified by the paper's epigraph); and second, that the chain of causation runs from irrational financial overreaction to real overinvestment (as seen in the quotation from the IMF report in the introduction). I propose a model which accounts for both perceptions, and yet suggests that they are fundamentally incomplete. The model implies that these bubble-like episodes are likely to accompany the emergence of new industries with particular characteristics: uncertainty about returns-to-scale and a competitive setting that protects production knowledge from free-riding. Given these conditions, predictable negative returns to financial claims and overshooting of real overinvestment can both result without the former having any role in determining the latter. Most importantly, the model does not include market failures, sub-optimal learning, or irrationality. While these are not incompatible with its mechanism, they do have strikingly different implications about the consequences of such episodes. In the completely frictionless case examined here, bubbles are actually the most efficient way to achieve the right long-run level of investment. If that is so, trying to prevent, regulate, or even identify them ex ante may be misguided.