Dynamic Regime Identification and Prediction Based on Observed Behavior in Electronic Marketplaces

We present a method for an autonomous agent to identify dominant market conditions, such as oversupply or scarcity. The characteristics of economic regimes are learned from historic data and used, together with real-time observable information, to identify the current market regime and to forecast market changes. The approach is validated with data from the Trading Agent Competition for Supply Chain Management. Introduction and Problem Definition The Supply-Chain Management Trading Agent Competition (Sadeh et al. 2003) (TAC SCM) involves a Supply Chain Management scenario in which in each round six autonomous agents attempt to maximize profit by selling personal computers they assemble from parts they buy from suppliers. The agent with the highest bank balance at the end of the game wins. An agent has to make many decisions, such as how many parts to buy, when to get them delivered, what types of computers to build, when to sell them, and at what price. Availability of parts and demand for computers varies randomly through the game and across market segments (low, medium, and high computer price). The market is affected not only by variations in supply and demand, but also by the actions of other agents. The small number of agents and their ability to adapt to and manipulate the market makes the game highly dynamic and uncertain. Marketing research methods have been developed for improving sales by understanding the relationships between market conditions and actions. For instance, in (Pauwels & Hanssens 2002), an analysis is presented on how in mature economics markets strategic windows of change alternate with long periods of stability. Sales strategies used in previous TAC SCM competitions model the probability of receiving an order for a given offer price, either by estimating the probability by linear interpolation from the minimum and maximum daily price records (Pardoe & Stone 2004), or by estimating the relationship between offer price and order probability with a linear cumulative density function (CDF) (Benisch et al. 2004), or by using a reverse CDF and factors such as quantity and due date (Ketter et al. 2004). Copyright c © 2005, American Association for Artificial Intelligence (www.aaai.org). All rights reserved. All these methods have the disadvantage that they do not explicitly take market conditions into account. The approach we present here is motivated by the longstanding realization among economists that market conditions play a critical role in making operational and strategic decisions for raw material acquisition, production, and sales. Regime Identification Mathematically, we define a regime with the help of a Gaussian mixture model (GMM). First we collect order prices from past games (in our experiments we use the semi-finals and finals of TAC SCM 2004). Each computer type has a different nominal price, which is simply the sum of the nominal cost of each component part. We use normalized prices (np) and we apply the Expectation Maximization algorithm to determine the components (μi,σi, and prior probability) of the GMM. The GMM has the following form: