Agent-Based and DSGE Macroeconomic Models: A Comparative Study

Lead Research Organisation: University of Surrey
Department Name: Economics


This project aims to compare and combine two major approaches to modelling: Dynamic Stochastic General Equilibrium (DSGE) and Agent-Based (AB) models. The former is arguably the best structured, most widely accepted, most sophisticated source of empirical information on the behaviour of national economies. The latter is arguably the most flexible tool for studying complex social systems with many dynamic and interacting components.

The two modelling approaches have different advantages and disadvantages. DSGE models may exaggerate individual rationality and foresight, and understate the importance of heterogeneity, that is, differences between agents focusing mostly on the way economic agents interact through aggregate prices. The AB models offer a more flexible approach to the role of other social interactions between individual agents in the economy by defining the characteristics and behaviour of individual heterogeneous agents with limited rationality, information and foresight. On the other hand, AB models may exaggerate errors in individual decision-making, since they usually model only simple strategies that are far from optimal choices and that evolve in time. The problem is that agents can depart from rationality in an infinite number of ways leading into what some economists refer to as a `wilderness'. The logical cohesion of rational expectations in DSGE models can be a benchmark for researchers interested in learning and bounded rationality.

This project will, we believe, be the first to compare the two modelling approaches systematically and draw conclusions on what each can learn from the other. Our general philosophy is to celebrate and exploit diversity in macroeconomic models. Our research strategy is to formulate AB model counterparts of DSGE models to reveal the relative strengths and limitations of the two modelling approaches. Our agentised DSGE (A-DSGE) models will provide important insights for strengthening the foundations of DSGE models while being better informed by economic theory than conventional AB models. Empirically, we will explore how AB models can benefit from the estimation approaches used in DSGE models, enabling us to perform a likelihood race between the traditional DSGE model and our agentised ones.

The project has three specific inter-related components of the research. First, we will take a conventional New Keynesian DSGE model as benchmark and build a DSGE model that more closely matches an AB model by relaxing the assumption of a representative agent with perfect information and cognitive abilities. This will involve multiple ways of modelling information limitations including the traditional approaches, as well as those associated with the `rational inattention' and `sticky information' literatures. We will investigate how the heterogeneity of agents can help to explain real world features and affect policy prescriptions.

Second, we will systematically compare our `best' DSGE model in terms of data fit with an AB macroeconomic model counterpart that has the same economic structure in terms of agents, markets and openness. We intend to draw upon existing methods of estimating DSGE models to estimate our ABM models, thus enabling us to perform a `likelihood race' (that maximizes the probability of observing selected macroeconomic data across the models on offer) between the traditional DSGE model and our agentized ones.

Finally, we will explore robust policies across these two contrasting modelling approaches. It is possible to design, for example, interest rate rules that are simple and robust across the rival models. Using this robust policy design methodology, our aim for macroeconomic policy recommendations is to avoid becoming `a prisoner of a single outlook' with respect to the modelling of expectations, departures from rationality in decision rules and aggregation.

Planned Impact

With the financial crisis has come widespread criticism of the apparent failure of standard economic models to predict the current troubles and their inability to explain what happened --- ranging from questions by the Queen to doubts expressed by the Deputy Governor of the Bank of England (see, for instance, The Economist (2010) Agents of change: Conventional economic models failed to foresee the financial crisis. Could agent-based modelling do better? 22nd July). Many of these criticisms were unfair or uninformed, but they did highlight the need not to be complacent and to consider alternative forms of macro-economic modelling or enhance existing approaches to make them more useful for understanding a world in recession. The proposed research aims to to do this, comparing and synthesising Dynamic Stochastic General Equilibrium (DSGE) models and Agent-Based (AB) modelling.

Hence the beneficiaries of the research are potentially all those who use macro-economic models for forecasting or for analysing the likely consequences of policies. This includes a wide range of organisations outside academia. For example, in the UK, the Treasury and departments including BIS, DEFRA, DWP and DFID are all concerned with the macro-economic implications of their policy decisions. In the private sector, the banks and other financial institutions use macro-economic models to some extent and are likely to use them more when their confidence in them is greater. The UK's expertise in macro-economic modelling is internationally renowned and improvements to methods of modelling will reverberate around the world, affecting government and private institutions that use macro-economic models everywhere.

Improvements in the ability of macro-economic models to represent actual economic processes and outcomes have the potential to benefit society through (a) helping to stabilise the economic system (b) allowing more effective planning of policy measures to, for example, reduce unemployment and increase wealth creation (c) allowing policies to move away from 'one size fits all' prescriptions to those that are appropriately personalised and sensitive to individual circumstances, by increasing the ability of economists to model social heterogeneity.

Such impacts are likely to become visible some five to ten years after the completion of the research, as this is typically the time taken for academic social science innovations to penetrate the public and private sectors. One of the most efficient ways of moving new ideas out of academia is through training students and researchers. The proposed project will employ two post-doctoral researchers who will become highly experienced in building agent-based models of the macro-economy. Two of the co-investigators (Hamill and Gilbert) are writing a textbook on agent-based modelling in economics for senior undergraduates and postgraduates, and the text will incorporate some lessons learned from the project. Finally both research centres will conduct self-financing training courses in DSGE and ABM methodologies.
Description Our achievements were: first, to synthesize Agent-Based (AB) and Dynamic Stochastic General Equilibrium (DSGE) approaches to macroeconomic modelling; second, to develop a new behavioural approach to construct and estimate macroeconomic models using experimental and macroeconomic data; third, to develop of a novel general methodology for the design of robust policy with model uncertainty and finally, to establish a network of researchers working on these and related issues.

The starting point of the research project was the criticism of the traditional DSGE paradigm with its strong assumptions of rational expectations (RE), a representative agent and market clearing. The framework neglects heterogeneity and the interaction between policy and inequality.

Much has been done within the DSGE community to address the above concerns and the first objective of three inter-related components of the research, to develop an extended general equilibrium paradigm (GE), was to contribute to this literature. Taking a conventional New Keynesian DSGE model as a benchmark, we built a series of DSGE models that more closely matches an agent-based (AB) model in some important respects namely, to relax the assumptions of a representative agent and perfect information. This involved multiple ways of modelling information limitations including the traditional approaches, as well as those associated with the `rational inattention' and `sticky information' literatures.

The second objective of the research was to systematically compare DSGE models with an AB macroeconomic model counterpart that has the same economic structure in terms of economic agents and markets. This study evolved into a contribution to a newly emerging literature in macroeconomics which we refer to as a Behavioural Framework (BF). The BF can be thought of as bridging the gap between DSGE and AB modelling in that rational expectations is replaced with learning, there is no assumption that forecast rules are well-specified (unlike the DSGE learning literature) and agents are heterogeneous in their beliefs and can learn from each other.

Taken together these two components led to the first two achievements stated above.

The final objective (third achievement) was to explore robust policies across different modelling approaches and thus avoid becoming `a prisoner of a single outlook' with respect to the modelling of expectations, departures from rationality in decision rules and aggregation. For example for monetary policy using the nominal interest rate as the instrument, the aim was to design an interest rate rule that is simple and robust across the rival models. To achieve this objective, we developed a novel general robust policy design methodology that is applicable to macro-economic policy in general, involving fiscal and macro-prudential alongside monetary policy.

This research has contributed to two network groups: the first one is coordinated by the Money, Macro and Finance Research Group which was short-listed for a full proposal for the ESRC Understanding the Macro-economy Network+; the second is the European Joint Doctoral Programme ExSIDE (Expectations and Social Influence Dynamics in Economics) which combines an interdisciplinary research agenda with an innovative European joint doctoral training programme.
Exploitation Route As mentioned in the previous section work is in progress with established collaborators to develop and apply the robust policy methodology to bounded and unbounded rationality. This is expected to be completed by the end of 2017.

We will exploit the two networks highlighted in the ``Significant Achievements'' Section above. First, we will link the new behavioural modelling approach more closely with experimental findings. This work will be carried out with Professor Cars Hommes, University of Amsterdam and Tinbergen Institute, who leads one of the world's foremost groups on experimental macroeconomics. Second, we will apply our behavioural macroeconomic framework to open advanced and emerging economies. This will exploit the Money, Macro and Finance Research Group network and an established collaboration between the School of Economics in the University of University and the Central Bank of Nigeria. Finally, we consider the systems estimation of new models to be of central importance. We are currently developing new methods of estimating highly non-linear behavioural models and we are working closely with Agent-Based modellers to apply our methods to their models. Both these projects and the input from experimental macroeconomics form part of the research agenda of the ExSIDE programme.
Sectors Financial Services, and Management Consultancy,Government, Democracy and Justice

Title Agent-Based Macroeconomic Demonstration Model 
Description This model is of a simple economy with markets for (i) a perishable consumer good, (ii) a capital good (like machinery and equipment) that depreciates (by time or usage) and is used for the production of the consumer good with and (iii) labour. The current version of the model is built as a basic framework for enabling explorative experiments under different scenarios (as explained below). The model combines various contexts and scenarios that may be relevant to representing an economy. Yet, this is still a demonstration model and the experiment results should not be taken too seriously because (i) the functional forms used in the model do not reflect well thought of assumptions or a good reflection of existing literature, they are rather crude functions chosen for easy identification of possible errors, the model doesn't incorporate (ii) growth in factors of production or advances in technology or supply side ("technology") shocks that are central in real business cycle theory. 
Type Of Material Computer model/algorithm 
Provided To Others? No  
Impact We are an interdisciplinary team of economists and sociologists. The impact at this stage is to facilitate a sharing of quite distinct modelling approaches. 
Description September 8-12, 2014, Third CIMS DSGE Modelling Summer School followed by a 1-day conference 
Form Of Engagement Activity A talk or presentation
Part Of Official Scheme? Yes
Geographic Reach International
Primary Audience Professional Practitioners
Results and Impact The School disseminated the DSGE modelling general methodology including the global solution methods being employed on the project ES/K005154/1.

PhD students, lecturers and researchers from many central banks and ministries attended this School. A result of the latter has been a stimulated interest in the use of DSGE models for policy-making.
Year(s) Of Engagement Activity 2014