昨日までの一連のエントリで紹介したサージェントインタビューでは学習理論が主なトピックの一つとなっていたが、インタビュアーの一人であるジョージ・エバンス(George Evans)が、最近イタリア国家統計局(Italian National Institute of Statistics=ISTAT)のMaurizio Boviがそのテーマについて書いた論文についてコメントした。それをオレゴン大学の同僚のMark Thomaが、Economist's Viewで――Boviの自論文紹介メールと共に――取り上げている


This is an interesting paper that has a lot of common ground with the adaptive learning literature. The techniques and a number of the arguments will be familiar to those of us who work in adaptive learning. The tenets of the adaptive learning approach can be summarized as follows: (1) Fully “rational expectations” (RE) are implausibly strong and implicitly ignores a coordination issue that arises because economic outcomes are affected by the expectations of firms and households (economic “agents”). (2) A more plausible view is that agents have bounded rationality with a degree of rationality comparable to economists themselves (the “cognitive consistency principle”). For example agents’ expectations might be based on statistical models that are revised and updated over time. On this approach we avoid assuming that agents are smarter than economists, but we also recognize that agents will not go on forever making systematic errors. (3) We should recognize that economic agents, like economists, do not agree on a single forecasting model. The economy is complex. Therefore, agents are likely to use misspecified models and to have heterogeneous expectations.

  1. 完全な「合理的期待」は非現実的なほど強力であり、経済的帰結が企業や家計(経済「主体」)の期待に影響されることによって生じる協調問題を暗に無視している。
  2. より現実的な見方は、主体が限定合理性を有しており、その合理性の程度は経済学者自身と同程度、というものである(「認知的整合性の原則」)。主体の期待が、時間を追って修正され更新される統計モデルに基づいている、というのがその例。この手法では、主体が経済学者より賢いという前提は回避されるが、同時に、主体が組織的な間違いを永遠に犯し続けることは無い、という認識を研究者は持っている。
  3. 経済主体は、経済学者と同様、単一の予測モデルで意見が一致することは無い、ということを我々は認識しなければならない。経済は複雑なものなのだ。従って、主体が誤ったモデルを使い、期待が均一なものとはならない、というのは良くあることである。


The focus of the adaptive learning literature has changed over time. The early focus was on whether agents using statistical learning rules would or would not eventually converge to RE, while the main emphasis now is on the ways in which adaptive learning can generate new dynamics, e.g. through discounting of older data and/or switching between forecasting models over time. I use the term “adaptive learning” broadly, to include, for example, the dynamic predictor selection literature.


Bovi’s paper “Are the Representative Agent’s Beliefs Based on Efficient Econometric Models” argues that with respect to GDP growth in the UK the answer to his question is no because 1) there is a single efficient econometric model, which is a version of AE (adaptive expectations), and 2) agents might be expected therefore to have learned to adopt this optimal forecasting model over time. However the degree of heterogeneity of expectations has not fallen over time, and thus agents are failing to learn to use the best forecasting model.
From the adaptive learning perspective, Bovi’s first result is intriguing, and merits further investigation, but his approach will look very familiar to those of us who work in adaptive learning. And the second point will surprise few of us: the extent of heterogeneous expectations is well-known, as is the fact that expectations remain persistently heterogeneous, and there is considerable work within adaptive learning that models this heterogeneity.

  1. 適応的期待の一バージョンである単一の効率的な計量経済モデルが存在する。
  2. 従って主体は、その最適な予測モデルに学習によってやがて適応するはずである。しかし実際には、期待の不均一性の程度は時間と共に減少することは無い。即ち、主体は最良の予測モデルを用いるように学習することに失敗している。