Market simulation comparasion

Introduction Project name Market simulation comparasion
Class: 4IT496 Simulation of Systems (WS 2012/2013)
Author: Jan Bečev
Model type: Agent-based simulation
Modeling software: NetLogo 5.0.1.

Problem definition

There is one market with more products on which we have standard buy and sell orders. Agent can change these orders. Point of simulation is to find how can one agent benefit from sentiment on market (so called bubble). There are more players on market, they got different possibility to be affected by short term changes on market (sentiment). Also these agents choose which product should they trade (from two possible). Point of simuation: Compare two models of market. A) Market where players are marginal (testing agents against real life data, which is not changed by our agents). Real life data will be taken from Patria for security ETFS Silver on London market (SLVR.L). B) Markets where players are mayor (market is result from actions of agents).

Method

Biggest quality of developer is doing task efficiently and reusing already done solutions of problem. During developement of model and simulation I have reused model of Multiagent simulation of market as a basic representation of market. For second simulation, where actions of agents do not have any result on market price, i have used exported data of observed market product (SLVR.L) for price making. I have used opening and closing prices for 150 consecutive days, therefore simulation has fix length of 300 steps. Fact, that our agents were not able to change price in second simulation, provided many benefits.
A) In fact, market contains so many agents and market value, that most of agents does not have any real power to change market price
B) We can see how good are those agents reacting in real enviroment and we can compare results to results in simulated enviroment, where agents have 100% price making power.
C) Simulation is less complex which brings benefit or results easier to understand.

Model

Agents

Agent types are from simulation „Fundamentalists and imitators“.

Fundamentalists

Marked white Decide to buy or sell product only based on internal value. Internal value is computed with function log-price.

Imitators

Marked green His action depends on action of others, either on surrounding 8 agents or agents from whole model. Outcome of previous turns increases or decreases chance of these agents to listen to others.

Imitators can have two states Optimist - violet Pesimist - black

Stubborns

Marked red Representing noisy traders, deciding at random whather buy or sell. These types of agents are result on using Multiagent market model as a basis of new model

Failed

Marked: yellow Lose shares, they stop involving other agents

Results

Main problem of model with Agent based price is it’s unstability. With basic configuration price usually escalates to one extrem and does not change this trend. Also there is problem, when we can’t try same simulation multiple times with same price. Interresting is fact, that more agents fail when simulation is run on agent-produced price than in price not determined by agents (in case when values of simulation are same).

Results with real data:

ALL fundamentalists

In case, when we deploy all agents like fundamentalists with no renew rate, there could be two outcomes. Either all agents fail (at step 63), or all survive. It is suprising, that this option (where agents have some, albeit limited, ability to asses situation on market) is often worse than random choice.

ALL Stubborns

In case when all agents are randomly choosing what to do, part of them fail but still some are able to survive.

ALL imitators

In case when all agents are deciding based on actions of surrounding agents, most agents fail, but still there are some able to survive .

Run with basic setting:

In most cases most agents fail, in rare cases most survive. In every run of simulaton at least some survived.

Runs with agents based price

ALL fundamentalists

In case, when we deploy all agents like fundamentalists with no renew rate in model with price determined by agents, none of agents fail.

ALL Stubborns

In case when all agents are randomly choosing what to do, part of them fail but still some are able to survive.

ALL imitators

In case when all agents are deciding based on actions of surrounding agents, none of agents fail.

Conclusion

Waste difference between results of agents when running in two different conditions showed problem of agent based simulations when applied to real life data. Problem lies in fact, that while model might seem to be logical, working and generating interesting results, in fact these results are not appliable to real life.
Another problem of model and market models in general, lied in fact that model would be either to simplified, or too complex. To balance need for simple model and model which would implement many different real life situations and difference of users, developers of basic model used many variables, which could be hardly translated to „real life“. It could be said, that it is problem only of this model, but this problem is very often with simulations which should bring some real-life implementations.
This could be described on a example when from simulation knows, that our ideal customer is customer with value beta between 0,25 and 0,5 and value gamma 0,87 and 0,9, while noone can describe what those values mean in real life.
Another problem of this simulation was poorly defined problem and required result. In one moment, problem was not described enought and simulation was required to be too complex so it was high above level of this course. Much easier and less complex simulations are presented as reserch papers with multiple authors.
Another problem with real life application of market simulations made in Netlogo is lack of basic support, documentation and functions implemented in language. For example lack of proper function for input of data from any other type of file than .txt in special format of data, which required change in file for data transfer from source (Patria).
Results of simulations were so different and so dependant on values, which did not had any representation in real-life that I would consider them pointless.

 

As a result, I would recommend for future simulations:

- Much less complex simulation topic, preferably defined by lecturer.
- Problem described in more detailed way, with defined results
- For simulations using real-life data or with expected real-life use program simulation in other tool than Netlogo because of its lacking imput functions and overall poor support
- Avoid topic of Markets because it either results in over simplified view, which brings useless results, or is too complex, which brings results not transferable to real life or is impossible to simulate.

Credits and References

Fundamentalists and imitators model by Marianna Caldana, Paolo Cova, Umberto Viano which can be found at http://web.econ.unito.it/terna/tesine/multiagent.htm