Extended Cognition in Economics Systems. Kevin McCabe Economics George Mason University

Transcription

1 Extended Cognition in Economics Systems by Kevin McCabe Economics George Mason University

2 Microeconomic Systems Approach See Hurwicz (1973) See Hurwicz and Reiter (2008) Environment: E g Performance: F Given (E, F, Z), choose (M, g, h), that realizes F on E. Outcomes: Z h Message Space: M

3 Environment Behavior Given the environment and institution what messages do people send? Experimental Economics See Smith (1982) Performance: Given the environment what are good outcomes Institution: Rules of Interaction Governs Messages Produces Outcomes Outcomes Message Space How do participants communicate with the institution and what feedback do they get.

4 Research Program on Spot Commodity Markets See Smith (1962) S Numeraire D Environment: (values/costs) Governance Institution: Double Auction Messages ID BID ASK ID B1 $3 B1 $4 $7 S3 $6 S1 Sell S Time Production Outcomes (contract Prices)

5 Computational Experimental Economics - Goals Tightly integrate, experimental economics with computational economics at the, algorithmic level, design level, code level, data level, and analysis level, to build better integrative research, and promote open-science in economics.

6 CEE - Approach Economics systems, are algorithms distributed in the environment, including agents, and in the institutions. These algorithms guide computations, by various platforms, including embodied brains.

7 CEE - Hypotheses Computational Experimental Economics Hypotheses, are on the algorithms, computations, and outcomes, produced by an economic system.

8 mtree Component Design Computational Agent Human Subject Controller Web Interface Environment Agent Agent Controller Agent Institution Controller Agent Agent Controller Controller Controller Institution Universe

9 mtree Component Interaction Design philosophy is to allow experimenters to implement simple and reusable code components Communication between each component is performed via message passing Experimenter need only design how to process messages based on directives and payload Message Contains: Sender, Receiver, Directive, Payload Agent Controller Processes Directives: Market Open Won auction, Bid/Ask

10 General Data Structure (DS) Control Variables (Fixed from level above) State Variables (Vary With Messages) Public (Available to Other Actors) Private (Unavailable) So pieces of the universe (institutions, environments, agents, can be easily shared by documenting message handling and data Structures.

11 Computational Experimental Economics Ecosystem Data Recorder Economic Universe Agents Subjects Monte Carlo Harness Experiment Controller

12 mtree Opensource at: gmucsn.github.io pip install mtree Contact us at:

13 CEE Example (Exchange in Spot Markets) Assume one goal of economic systems design 1 is to facilitate exchange by minimizing transactions costs. One way to measure costs are in terms of the costs of running the distributed algorithms that produce exchange. 1 Note design could be constructive design or spontaneous design. See Vernon Smith, Rationality in Economics, 2008.

14 CEE Zero-Intelligence-Traders (Budget Constrained) If these costs are concentrated in those algorithms running on embodied brain platforms them we may reasonable expect that economic systems will organize to minimize these costs Budget constrained ZI implementation: Buyer i s bid for unit j is a random draw from [0,v ij ], where v ij,is i s valuation for the j th unit. Seller i s ask for unit j is a random draw from [c ij, B], where c ij is j s cost of producing the j th unit. 2 See Gode and Sunder (1993) and follow up research.

15 CEE Zero-Intelligence-Traders (Basic Results) Supply and Demand ZI Traders with Budget Constraint Human Traders Period

16 CEE Zero-Intelligence-Traders (Why do they work) Cason and Friedman (1993). Trades with greater exchange surplus tend to trade earlier in any period. Trade info added

17 CEE Zero-Intelligence-Traders (Why do they work) Due to the improvement rule in the DA, Zero Intelligence Traders tend to close the bid-ask spread in favor of higher value or lower cost units. Standing Ask Standing bid

18 CEE Zero-Intelligence-Traders (When do they fail) Brewer, Nelson, and Plott (2002) continuously refresh units of supply and demand as units are trade. Human subjects still do fairly well, but ZI traders fail badly. Humans ZI Traders

19 CEE Zero-Intelligence-Traders (Adding Learning) Dave Cliff (1997) Suggests a model of ZIP traders who adjust profit margins. If you are a buyer you can earn, e = V p. Suppose you want a profit margin, m between [0, 1], then mv = V p, or p = (1-m)V. You then bid (1-m)V. Profit margins can then be moved up and down based on the success of this strategy. Cliff uses the last accepted bid or ask to update m.

20 CEE Double Auction Strategies A strategy for player, i, is a mapping from i s information set I i to a non-negative real valued bid or ask, that is, s i : I i {Bid, Ask} x R + The information set I i is a subset of I, where I = {order flow, standing BID/ASK, contracts, time remaining in current period, values, costs, #Buyers, #Sellers, initial BID/ASK, auction disposition, order processing {rejected, standing, contract}, contract processing {buyer accepted, seller accepted}}

21 CEE Double Auction Tournament It turns out, except for some very simple environments it has been impossible to characterize the optimal Nash Equilibrium strategy for the one unit spot double auction. Rust, Miller, Palmer (1992) report on trading strategies used in the Santa Fe Institute Double Auction Tournament which offered $10,000 in prize money. Top two strategies were snipers that simply took the standing bid or ask when the bid-ask spread got within a certain range. Obviously a sniper will not work if there are no market makers.

22 CEE What are we doing now We have built a single unit double auction system. Currently this is being implemented in mtree. Undergraduates are taking experimental papers from the literature and building them in the double auction universe, and running simulations exploring the various strategies described above. We are building an Agent Zero (Joshua Epstein, 2013) neuroeconomics inspired agent for these games.

23 Thank You

24 Economic Universe {DS} Environment {DS} Institution 1 {DS} Institution 2 {DS} Institution 3 {DS}

25 Economic System{DS} Environment {DS} Institution {DS} Agent 1 {DS} Agent k {DS} Agent N {DS}

26 Micro Economic System (One Institution Framework) Run Environment(Control Variables) Connect Agent K Initialize Agent Start Institution Respond to Get and Change state Requests Run Institution (Control Variables)Reve Open Institution Respond to Get state messages Change state based on messages If state == Public: Inform Environment of state changes

27 Spot Market System Run Environment(supply and demand) Connect Agent K Initialize Agent K as Buyer or Seller Start Double Auction (Change Status) Get contract state change Run One Unit Double Auction Without Book Open Auction Respond to Get Standing BID and ASK msg Change state based on BIDS and ASKS If contract: Inform Environment of Contract

28 Spot Market Agent Get Values or Costs from Environment Wait for Double Auction to Open Initialize other state variables my_offers = {}; cur_unit = 0 While Double Auction is Open: Get Contracts If new_contract: update state Get Standing Bid and Ask Decide on new_offer (Bid/buyer; Ask/Seller) if new_offer: send new_offer

29 Spot Market Algorithm Consists of an initial condition (control variables) Followed by a sequence of get messages from the Environment, Institution(s), and Agent(s) that change the internal states of these actors and lead to decisions and change messages which further change the state of the actors.

30 Spot Market Performance Consists of an initial condition (control variables) Followed by a sequence of get messages from the Environment, Institution(s), and Agent(s) that change the internal states of these actors and lead to decisions and change messages which further change the state of the actors.