Fundamentalists vs. chartists: Learning and predictor choice dynamics

In a simple, forward looking univariate model of price determination we investigate the evolution of expectations dynamics in the presence of two types of agents: fundamentalists and chartists. In particular, we combine evolutionary selection among heterogeneous classes of models through predictor choice dynamics based on a logit model, with adaptive learning in the form of parameters updating within each class of rules. We find that, for different parameterizations, it can happen that fundamentalists drive chartists completely out of the market or vice versa, and also that heterogeneous equilibria in which fundamentalists and chartists coexist are possible. Interestingly, though, only equilibria in which fundamentalists outperform chartists turn out to be adaptively learnable by agents.     

Forecasting ATM cash demands using a local learning model of cerebellar associative memory network

Forecasting cash demands at automatic teller machines (ATMs) is challenging, due to the heteroskedastic nature of such time series. Conventional global learning computational intelligence (CI) models, with their generalized learning behaviors, may not capture the complex dynamics and time-varying characteristics of such real-life time series data efficiently. In this paper, we propose to use a novel local learning model of the pseudo self-evolving cerebellar model articulation controller (PSECMAC) associative memory network to produce accurate forecasts of ATM cash demands. As a computational model of the human cerebellum, our model can incorporate local learning to effectively model the complex dynamics of heteroskedastic time series. We evaluated the forecasting performance of our PSECMAC model against the performances of current established CI and regression models using the NN5 competition dataset of 111 empirical daily ATM cash withdrawal series. The evaluation results show that the forecasting capability of our PSECMAC model exceeds that of the benchmark local and global-learning based models.     

Learning in an estimated medium-scale DSGE model

We evaluate the empirical relevance of learning by private agents in an estimated medium-scale DSGE model. We replace the standard rational expectations assumption in the Smets and Wouters (2007) model by a constant-gain learning mechanism. If agents know the correct structure of the model and only learn about the parameters, both expectation mechanisms produce very similar results, and only the transition dynamics that are generated by specific initial beliefs seem to improve the fit. If, instead, agents use only a reduced information set in forming the perceived law of motion, the implied model dynamics change and, depending on the specification of the initial beliefs, the marginal likelihood of the model can improve significantly. These best-fitting models add additional persistence to the dynamics and this reduces the gap between the IRFs of the DSGE model and the more data-driven DSGE-VAR model. However, the learning dynamics do not systematically alter the estimated structural parameters related to the nominal and real frictions in the DSGE model.     

Escape dynamics: A continuous-time approximation

We extend a continuous-time approximation approach to the analysis of escape dynamics in economic models with constant gain adaptive learning. This approach is based on the application of the results of continuous-time version of large deviations theory to the linear diffusion approximation of the original discrete-time dynamics under learning. We characterize escape dynamics by analytically deriving the most probable escape point and mean escape time. The approximation is tested on the Phelps problem of a government controlling inflation while adaptively learning a misspecified Phillips curve, studied previously by Sargent (1999) and Cho et al. (2002) (henceforth, CWS), among others. We compare our results with simulations extended to very low values of the constant gain and show that, for the lowest gains, our approach approximates simulations relatively well. We express reservations regarding the applicability of any approach based on large deviations theory to characterizing escape dynamics for economically plausible values of constant gain in the model of CWS when escapes are not rare. We show that for these values of the gain it is possible to derive first passage times for learning dynamics reduced to one dimension without resort to large deviations theory. This procedure delivers mean escape time results that fit the simulations closely. We explain inapplicability of large deviations theory by insufficient averaging near the point of self-confirming equilibrium for relatively large gains which makes escapes relatively frequent, suggest the changes which might help approaches based on the theory to work better in this gain interval, and describe a simple heuristic method for determining the range of constant gain values for which large deviations theory could be applicable.     

Simulation‐based tests of forward‐looking models under VAR learning dynamics

In this paper we propose a simulation‐based technique to investigate the finite sample performance of likelihood ratio (LR) tests for the nonlinear restrictions that arise when a class of forward‐looking (FL) models typically used in monetary policy analysis is evaluated with vector autoregressive (VAR) models. We consider ‘one‐shot’ tests to evaluate the FL model under the rational expectations hypothesis and sequences of tests obtained under the adaptive learning hypothesis. The analysis is based on a comparison between the unrestricted and restricted VAR likelihoods, and the p‐values associated with the LR test statistics are computed by Monte Carlo simulation. We also address the case where the variables of the FL model can be approximated as non‐stationary cointegrated processes. Application to the ‘hybrid’ New Keynesian Phillips Curve (NKPC) in the euro area shows that (i) the forward‐looking component of inflation dynamics is much larger than the backward‐looking component and (ii) the sequence of restrictions implied by the cointegrated NKPC under learning dynamics is not rejected over the monitoring period 1984–2005. Copyright © 2010 John Wiley & Sons, Ltd.     

Distribution of the maximum of the number of impulses at any instant in a type II counter in a given interval of time

Summary Here we have considered a type II counter in which the inter-arrival times of particles are independent identically distributed random variables following a given distribution. Each particle is assumed to give rise to an impulse and the state of the counter at an instant is defined by the number of impulses present at that instant. Two variations of the model are taken into account depending upon whether the system starts initially with an impulse or not. We have obtained integral equations for the distribution of the maximum of the number of impulses at any instant in a given interval of time for the two models considered. Special cases where the inter-arrival times and the duration of the impulses follow the exponential distribution have also been dealt with. Dedicated to Professor V.Ganapathy Iyer on the occasion of his sixty first birth day.     

Statistical learning and exchange rate forecasting

This study uses innovative tools recently proposed in the statistical learning literature to assess the capability of standard exchange rate models to predict the exchange rate in the short and long runs. Our results show that statistical learning methods deliver remarkably good performance, outperforming the random walk in forecasting the exchange rate at different forecasting horizons, with the exception of the very short term (a period of one to two months). These results were robust across countries, time, and models. We then used these tools to compare the predictive capabilities of different exchange rate models and model specifications, and found that sticky price versions of the monetary model with an error correction specification delivered the best performance. We also explain the operation of the statistical learning models by developing measures of variable importance and analyzing the kind of relationship that links each variable with the outcome. This gives us a better understanding of the relationship between the exchange rate and economic fundamentals, which appears complex and characterized by strong non-linearities.     

Trader matching and the selection of market institutions

We analyze a stochastic dynamic learning model with boundedly rational traders who can choose among trading institutions with different matching characteristics. The framework allows for institutions featuring multiple prices (per good), thus violating the “law of one price.” We find that centralized institutions are stochastically stable for a broad class of dynamics and behavioral rules, independently of which other institutions are available. However, some decentralized institutions featuring multiple prices can also survive in the long run, depending on specific characteristics of the underlying learning dynamics such as fast transitions or optimistic behavior.     

Econometric models of asymmetric ascending auctions

We develop econometric models of ascending (English) auctions which allow for both bidder asymmetries as well as common and/or private value components in bidders’ underlying valuations. We show that the equilibrium inverse bid functions in each round of the auction are implicitly defined (pointwise) by a system of nonlinear equations, so that conditions for the existence and uniqueness of an increasing-strategy equilibrium are essentially identical to those which ensure a unique and increasing solution to the system of equations. We exploit the computational tractability of this characterization in order to develop an econometric model, thus extending the literature on structural estimation of auction models. Finally, an empirical example illustrates how equilibrium learning affects bidding during the course of the auction.     

Forecasting U.S. money growth using economic uncertainty measures and regularisation techniques

This paper examines the out-of-sample forecasting properties of six different economic uncertainty variables for the growth of the real M2 and real M4 Divisia money series for the U.S. using monthly data. The core contention is that information on economic uncertainty improves the forecasting accuracy. We estimate vector autoregressive models using the iterated rolling-window forecasting scheme, in combination with modern regularisation techniques from the field of machine learning. Applying the Hansen-Lunde-Nason model confidence set approach under two different loss functions reveals strong evidence that uncertainty variables that are related to financial markets, the state of the macroeconomy or economic policy provide additional informational content when forecasting monetary dynamics. The use of regularisation techniques improves the forecast accuracy substantially.     

Learning and coordination with dispersed information

We analyse adaptive learning in a model of incomplete and dispersed information, with externalities and strategic interactions. We build on the framework proposed by Angeletos and Pavan (2007a) and extend it to a dynamic multi-period setting where agents need to learn to coordinate. We derive conditions under which adaptive learning obtains in such setting and show that, when actions are strategic substitutes, the information structure affects the speed of convergence: while more precise private information is beneficial, better public information has negative effects. We also show that adaptive learning dynamics converge to the Bayesian Nash equilibrium, which means that agents can learn to act strategically by relying only on observable (exogenous) information.     

Discrete Dynamics for the Core-periphery Model

Abstract

It is known that the discretisation of continuous-time models can introduce chaotic behaviour, even when this is not consistent with observations or even the model's assumptions. We propose generic dynamics describing discrete-time core-periphery models that comply with the established assumptions in the literature and are consistent with observed behaviour. The desired properties of the dynamics are proved analytically in the general case. We also give particular forms for the dynamics for those interested in applying our model.     

Application of collective knowledge diffusion in a social network environment

ABSTRACT

We present a group dynamics model that shows knowledge integration as a process occurring over time. As each individual in the group contact others, his own knowledge changes, and over time the collective knowledge is obtained. This allows modeling knowledge diffusion in a social network and while the models presented in this paper are not competitive in that area, they approach the problem from previously unconsidered direction. We test the behavior of the model in a multi-agent simulation and we test a simple advertisement campaign in a social network. We provide discussion of elements needed for making model more competitive.     

A General Approach to Testing Volatility Models in Time Series

Volatility models have been playing important roles in economics and finance. Using a generalized spectral second order derivative approach, we propose a new class of generally applicable omnibus tests for the adequacy of linear and nonlinear volatility models. Our tests have a convenient asymptotic null N(0,1) distribution, and can detect a wide range of misspecifications for volatility dynamics, including both neglected linear and nonlinear volatility dynamics. Distinct from the existing diagnostic tests for volatility models, our tests are robust to time-varying higher order moments of unknown form (e.g., time-varying skewness and kurtosis). They check a large number of lags and are therefore expected to be powerful against neglected volatility dynamics that occurs at higher order lags or display long memory properties. Despite using a large number of lags, our tests do not suffer much from the loss of a large number of degrees of freedom, because our approach naturally discounts higher order lags, which is consistent with the stylized fact that economic or financial markets are affected more by the recent past events than by the remote past events. No specific estimation method is required, and parameter estimation uncertainty has no impact on the convenient limit N(0,1) distribution of the test statistics. Moreover, there is no need to formulate an alternative volatility model, and only estimated standardized residuals are needed to implement our tests. We do not have to calculate tedious and model-specific score functions or derivatives of volatility models with respect to estimated parameters, which are required in some existing popular diagnostic tests for volatility models. We examine the finite sample performance of the proposed tests. It is documented that the new tests are rather powerful in detecting neglected nonlinear volatility dynamics which the existing tests can easily miss. They are useful diagnostic tools for practitioners when modelling volatility dynamics.     

Option pricing with discrete time jump processes

In this paper we propose new option pricing models based on class of models with jumps contained in the Lévy-type based models (NIG-Lévy, Schoutens, 2003, Merton-jump, Merton, 1976 and Duan based model, Duan et al., 2007). By combining these different classes of models with several volatility dynamics of the GARCH type, we aim at taking into account the dynamics of financial returns in a realistic way. The associated risk neutral dynamics of the time series models is obtained through two different specifications for the pricing kernel: we provide a characterization of the change in the probability measure using the Esscher transform and the Minimal Entropy Martingale Measure. We finally assess empirically the performance of this modelling approach, using a dataset of European options based on the S&P 500 and on the CAC 40 indices. Our results show that models involving jumps and a time varying volatility provide realistic pricing and hedging results for options with different kinds of time to maturities and moneyness. These results are supportive of the idea that a realistic time series model can provide realistic option prices making the approach developed here interesting to price options when option markets are illiquid or when such markets simply do not exist.     

Equivalent representations of non-exponential discounting models

I characterize the entire class of consumption rules for finite-horizon models in which consumption is proportional to lifetime wealth. Any such rule can be obtained from a preference model with CRRA period utility. In a steady state with constant interest rates, a proportional consumption rule can be derived from a model with time-consistent preferences or from a model with possibly time-inconsistent preferences in which a household continually reoptimizes future utility discounted relative to the present instant. These two preference models will only coincide for the special case when the discount function is exponential. More generally, there will be two distinct yet observationally equivalent preference models. Hyperbolic-like discounting may arise because that is a simpler way for the brain to process a standard exponential discount function after accounting for mortality risk.     

Strategic learning and entry-equilibrium

This paper describes the nature of equilibrium for a dynamic game between a monopolist and an entrant who adopts a learning strategy for cost reduction. The learning process is characterized as a continuous-time Poisson process where each success reduces a part of the entrant's cost difference with the incumbent. The entrant chooses a stream of expenditure after every success to control the instant probability of the next success. The techniques of semi-Markov decision processes are used to characterize equilibrium learning decisions at every stage, distribution of entry times, and the effects of subsidies, technical progress, etc. on pre-entry and post-entry learning strategies.     

-learning agents in a Cournot oligopoly model

Q-learning is a reinforcement learning model from the field of artificial intelligence. We study the use of Q-learning for modeling the learning behavior of firms in repeated Cournot oligopoly games. Based on computer simulations, we show that Q-learning firms generally learn to collude with each other, although full collusion usually does not emerge. We also present some analytical results. These results provide insight into the underlying mechanism that causes collusive behavior to emerge. Q-learning is one of the few learning models available that can explain the emergence of collusive behavior in settings in which there is no punishment mechanism and no possibility for explicit communication between firms.     

A component model for dynamic correlations

We propose a model of dynamic correlations with a short- and long-run component specification, by extending the idea of component models for volatility. We call this class of models DCC-MIDAS. The key ingredients are the Engle (2002) DCC model, the Engle and Lee (1999) component GARCH model replacing the original DCC dynamics with a component specification and the Engle et al. (2006) GARCH-MIDAS specification that allows us to extract a long-run correlation component via mixed data sampling. We provide a comprehensive econometric analysis of the new class of models, and provide extensive empirical evidence that supports the model’s specification.     

Learning to bid in sequential Dutch auctions

We propose an agent-based computational model to investigate sequential Dutch auctions with particular emphasis on markets for perishable goods and we take as an example wholesale fish markets. Buyers in these markets sell the fish they purchase on a retail market. The paper provides an original model of boundedly rational behavior for wholesale buyers׳ behavior incorporating learning to improve profits, conjectures as to the bids that will be made and fictitious learning. We analyze the dynamics of the aggregate price under different market conditions in order to explain the emergence of market price patterns such as the well-known declining price paradox and the empirically observed fact that the very last transactions in the day may be at a higher price. The proposed behavioral model provides alternative explanations for market price dynamics to those which depend on standard hypotheses such as diminishing marginal profits. Furthermore, agents learn the option value of having the possibility of bidding in later rounds. When confronted with random buyers, such as occasional participants or new entrants, they learn to bid in the optimal way without being conscious of the strategies of the other buyers. When faced with other buyers who are also learning their behavior still displays some of the characteristics learned in the simpler case even though the problem is not analytically tractable.