You are on page 1of 1

A stochastic, agent-based, actioned by heterogeneous adaptivity

(SABAHA) model for speculative bubbles.

Duc Pham-Hi,
ECE Graduate School of Engineering,
Paris France

EAEPE2015 abstract: This paper reports an effort to model speculative bubble formation in economies of
developing countries where (a) a large shadow credit sector exists, e.g. from foreign investors (b) a
widespread, diffuse and omnipresent population of small entrepreneurs act both as consumers and
producers and amplify speculations, and (c) monetary authorities and the behaviors of the institutional
lenders are not as precautionary as in the larger western financial places.

In so doing, it will also outline a methodology suitable for investigating out-of-equilibrium chain of
events.

First, 3 basic elements of the dynamic stochastic general equilibrium framework are shortly re-examined:
the use of Euler-Lagrange related optimization in structurally evolving dynamics, the dependence on a
tether to a steady state around which perturbations are studied, and finally the meaning itself of the
optimization which encompass all the actors of the economy. These points of discussion clear the way
for the introduction of alternative mechanisms we propose at the heart of a new model.

Next, the architecture of the model presented: stochastic, agent-based, actioned by heterogeneous
adaptivity (SABAHA). The heterogeneous agents - aggregate consumer, producer, banking and credit
system, and government, in the base version - have each their own different prediction mechanisms and
optimization processes. There is no unique, instantaneous, across-markets, optimization process leading
to a best Paretian compromise for all. Each agent may err in his personal forecast, leading to losses/gains
when the other agents reveal their control variables sets, and will adapt his arbitrage between
preferences for the present vs. of the future using a reinforcement learning mechanism. This preference
present/future provides the basis for online sampling and optimization attached to each individual
agent.

A section will be devoted to the algorithms that allow prediction by agents. They are based on non-
Gaussian distributions, and rely on Sequential Monte Carlo techniques. This mechanism avoids the
Markov hypothesis for the transition between states, and it also allow the forking of scenarios thanks to
the multimodal distributions.

Similarly, a section will deal with the optimization process which now breaks away from Euler-Lagrange.
Reinforced learning theory provides the core tools in this part.

Finally, remarkable sample paths from this model will be extracted, and analyzed, to show how far away
from an equilibrium state the agents can be led, by the growing gaps between their expectations and
observations. This framework is currently under effort to be fitted to model bubbles in events in Vietnam
during the 2007-2012 period. Observations of prominent real estate bubbles from imported foreign
capital, 2-digit inflation, State Bank of Vietnam interventions and evidence of a shadow sector can serve
as example.

You might also like