Announcements
New international banking rules would not prevent another financial crisis
The Basel III regulatory framework, as planned, will not reduce systemic risk in the financial sector, according to new research. Instead, regulations should aim to increase the resilience of financial networks. Current regulations aimed at reducing risk of crisis in the financial sector will not effectively reduce that risk, according to new research from the International Institute for Applied Systems Analysis (IIASA), published in the Journal of Economic Dynamics and Control. Introducing regulations that aim to increase the system network resilience would be more effective, the study shows. The Basel III framework is a new set of international banking regulations that were proposed after the financial crisis in 200708, with the goal of reducing the risk of a future banking crises. The regulations, which are currently under intense discussion, would set higher requirements for bank capital and liquidity reserves and introduce capital surcharges for systemically important banks—those that are “too big to fail.” The aim of the Basel III regulations is to reduce the risk of systemwide shocks in the financial sector. However, the study shows that the capital surcharges would have to be much higher that currently set in order to be effective, and that would lead to a severe loss of efficiency in the financial system. “The recent financial crisis clearly indicates that a resilient banking sector in terms of underlying financial networks is a necessary condition for achieving sustained economic growth. It is therefore essential that Basel III, the upcoming international regulatory framework for banks, really address the problem of systemic risk in the financial system,” says IIASA researcher Sebastian Poledna, who led the study. The research is based on a stateoftheart agentbased model of a financial system and the real economy. In particular, the study focused on banks that are “too big to fail,” known as globally systemically important banks (GSIBs). Using the model, the researchers ran a series of experiments simulating different types of regulations and their impacts on the financial system risk and resilience. Replacing the currently proposed Basel III regulations with different regulations that aim to restructure financial networks would be much more effective in increasing resilience while avoiding the loss of efficiency in markets, according to the study. Such methods could include smart transaction taxes based on the level of systemic risk, which the researchers proposed in a recent study, in order to reshape the topology of financial networks. "The new regulation scheme Basel III claims to explicitly address systemic risk. We were surprised to find how little it really does so under realistic scenarios. The study highlights how important datadriven agentbased modeling has become as a tool to help us identify unintended consequences of regulations and propose more effective solutions,” says IIASA researcher Stefan Thurner, who coauthored the study. “The international banking system is complex and intricately connected,” explains Poledna. “In order to make intelligent regulations, it’s important to analyze how regulations will affect financial networks from a systemic perspective.” Reference Poledna S, Bochmann O, & Thurner S (2017). Basel III capital surcharges for GSIBs are far less effective in managing systemic risk in comparison to networkbased, systemic riskdependent financial transaction taxes. Journal of Economic Dynamics and Control 77: 230246. DOI:10.1016/j.jedc.2017.02.004. Katherine Leitzell Presse und ÖffentlichkeitsarbeitInternational Institute for Applied Systems Analysis (IIASA) 
Bayesian analysis tutorial using PyMC
This is a short Bayesian analysis tutorial developed around the following problem: Consider the following dataset, which is a time series of recorded coal mining disasters in the UK from 1851 to 1962 [Jarrett, 1979]. The first step in the analysis is the Bayesian model building. We assume that the occurrences of the disasters can be modelled as a poisson process. Our hypothesis is, that at some point in time the mining process switched to a new technology, resulting in a lower rate of disasters in the later part of the time series. We define the following variables for the Bayesian stochastic model:
The model can then be defined as where D is dependent on s, e, and l. In a Bayesian network the probabilistic variables s, e, and l are considered as 'parent' nodes od D and D is a 'child' node of s, e, and l. Similarly, the 'parents' of s are t_l and t_h, and s is the 'child' of t_l and t_h.
The nest step is fitting the probability model (linked collection of probabilistic variable) to the recorded mining disaster time series. This means we are trying to represent the posterior distribution. Markov chain Monte Carlo (MCMC) algorithm [Gamerman 1997] is the method of choice. In this case we represent the posterior p(s,e,lD) by a set of joint samples of it. The MCMC sampler produces these samples by randomly updating the values of s, e and l for a number of iterations. This updating algorithm is called MetropolisHasting [Gelman et al. 2004]. With a reasonable large number of samples, the MCMC distribution of s, e and l converges to the stationary distribution. There are many generalpurpose MCMC packages. Here I use PyMC  a python module created by David Huard and Anand Patil and Chris Fonnesbeck. I prefer a programming interface to stand alone programs like WinBugs for its flexibility. It uses high performance numerical libraries like numpy and optimized Fortran routines. The following code imports the model, instantiates the MCMC object and run the sampler algorithm: >>> from pymc.examples import DisasterModel >>> from pymc import MCMC >>> M = MCMC(DisasterModel) >>> M.isample(iter=10000, burn=1000, thin=10) Below are the 900 samples (left) of variable s  the year in which the rate parameter changed. The histogram (right) with a mean at 40 and 95% HPD interval (35, 44). Next, the 900 samples (left) of variable e  the early rate parameter prior to s. The histogram (right) with a mean at 3.0 and 95% HPD interval (2.5, 3.7). Finally, the 900 samples (left) of variable l  the late rate parameter posterior to s. The histogram (right) with a mean at 0.9 and 95% HPD interval (0.7, 1.2).

Econometric Multiplayer Game
In order to harvest behavioral patterns of market participants Doyne Farmer came up with the idea to use an agent based economic model in a gaming mode. The idea is to replace some or all of the artificial agents by real people. In a "wargaming style" user of a multiplayer game will be able to develop an intuition for consequences of individual economic decisions in a dynamic interconnected gaming environment. The data can be used to understand human decision making. Behavioral information can be useful in calibrating econometric agent based models. Clearly, it will be quite challenging to design a game engaging enough to draw the attention of the gaming community. Also there are serious doubts that the data can be used for a real world model, because behavioral patterns in a game environment seem to be quite different to behavior in reality. Nevertheless, I think the idea is interesting enough to give it a try. Below I will outline some architectural considerations of such a game. If you are just interested in playing the game you can go right here: http://game.ComplexLab.org. Keep in mind that this is a first prototype. It will be evolving constantly and there is no guaranty of service. Architecture of Multiplayer GamesAt some point of the game the outcome of individual actions need to be synchronized among different computers/processes. In general, there are two common architectures for resolving arbitration: (1) clientserver and (2) peertopeer. Clientserver is conceptual simpler, easier to implement and better suited to generate game data. ClientServerEach user of the game and each artificial agent runs a local client program. The client programs are connected to a central machine  the server program. The server program is maintaining the state of the game and is broadcasting this information to the individual clients. This design makes the server the bottleneck, both computational and bandwidthwise and it may turn out to be a serious scaling problem. On the other hand it is easy to maintain game state and access control. The minimal example of a client program is a terminal. It transmits user inputs to the server and reports server messages to the user. The main loop of the client program would look like this:
Realtime Push NotificationAjax Push Engine (APE) is an open source realtime push notification system for streaming data to any browser using web standards only (Asynchronous JavaScript and XML (AJAX)). It includes a comet server and a Javascript Framework. Here is a demo that shows how APE can handle massive multiuser moving on a web page in realtime.PeertoPeer In a peartopear system (P2P) each user of the game runs the same peer program, or at least groups of users run the same program. The peer program maintains the local state e.g. the position of the player. When moving, the peer program is also responsible for collision avoidance. Therefore, peers need to broadcast their state to other peers. A minimalistic example of a peer loop would look like this:
There are several issues with this type of architecture that need to be adressed:
Possible Gaming ScenariosFarmer at all [1] proposed a number of gaming scenarios:
To be continued. 
Computational Mechanics
Computational Mechanics is an extension of the approaches typically found in statistical mechanics. It is a technique to describe both temporal and spacial organization in systems. It is a more detailed structural analysis of behavior than those captured solely in terms of probability and degrees of randomness. It complements and augments more traditional techniques in statistical physics by describing the structural regularies inherent in the system or data. Beyond focusing on measures of disorder, computational mechanics aims to detect and quantify structure in natural processes. Natural information processing mechanisms can be represented and analyzed in a diverse set of model classes. Contemporary notions of computation and of useful information processing must be extended in order to be useful within natural processes and their instantiation in multiagent systems. This is because natural processes are systems that can be spatially extended, continuous, stochastic, or even some combination of these and other characteristics fall often outside the scope of discrete computation theory. Computational mechanics attempts to construct the minimal model capable of statistically reproducing the observed data. This is called an εmachine or causal state machine. This approach leads one naturally to address the issue of the information processing capability of the system. There is an vast amount of literature on computational mechanics. This includes both, theoretical developments as well as system applications. Jim Crutchfield maintains a rather exhaustive page on computational mechanics called the Computational Mechanics Archive, and almost all papers of interest are listed there. εmachine (causal state machine)The εmachine is a mathematical object that embodies the Occam's Razor principle in that its description of the process has minimal statistical complexity subject to the constraint of maximally accurate prediction. The εmachine provides
εmachine reconstruction The problem is inferring εmachines from limited data is called εmachine reconstruction. The focus is on developing inference algorithms (statemerging versus statecreating) and accompanying statistical error theories, indicatiing how much data and computing time is required to attain a given level of confidence in an εmachine with an finite number of causal states.causalityCausality is defined in temporal terms: Something causes something else by preceding it. A sequence of causal states has the property that causal states at different times are independent, conditioned on the intermediate causal states. This is a Markov chain: In knowing the whole history of the process up to a point has no more predictive capacity than knowing the causal state it's in at that point. The causal structure is the equivalence relation causal states establish among different histories which have the same conditional distribution of future observables and the symbol emitting state transition probability. A directed graph where states are represented as nodes and edges labeled with emitted symbols and transition probabilities represent stete transitions can be used to visualize the causal information in the εmachine.type of raw data for inferring causality (suitable processes)Processes best captured by this method are sequences of a symbolic logic character, with complex pattern and an unknown generative mechanism, such as an extremelylong string on nottotallyrandom 1's and 0's. Processes are assumed to be stationary, however online reconstruction can relax this constraint. Algorithmic complexity of the reconstruction increases exponentially with the size of the symbol set. The even process is reconstructed (using CSSR) from sequences of 10000 samples with correct causal states and transition probabilities within 3% of the correct probabilities. The even process is a process which generates a string of 1's and 0's according to the rule that after a 0 is emitted, there is a coin flip to see whether a 1 or 0 follows. After an odd number of 1's being emitted, the process must emit a 1, but after an even number of 1's it is back to the coin flip. Randomness is embodied in the minimally stochastic transitions form one state to another. The given state can lead to many states, but a given symbol leads to at most one of those states.limitationsUnlike for graphical causal models a "factor" is not a meaningful term for the definition of an εmachine. In computational mechanics, the causal state definition sidesteps this, because it doesn't matter what determines the causal state to be a causal state, it just depends on the future morph of a given history.Pearl introduces the notion of "stability" (robustness of independence patterns) is not satisfied in εmachine reconstruction. This is when the model's structure can retain its independence pattern in the face of changing parameters. It is unclear in this context what the "parameters" actually are.

AgentBased Model of the Housing Bubble
This article outlines an approach in agentbased computational economics to build a macroeconomic model of the recent housing bubble. The goal of this effort is to gain a better understanding of it's causes and to formulate policy prescriptions. The common view of economists with respect to the housing bubble causes is that the Federal Reserve policies and measures that started in 1993 with respect to interest rates has led to the current crisis. This hypothesis needs to be verified (by observing some emergent properties). This is a model where we observe patterns of development out of the individual interactions of agents qualitatively and try to match them with the empirical data. In contrast to the conventional approach that produces quantitative house price projections based on trends in incomes, interest rates and housing supply and demand, the agentbased approach simulates the interaction of individual agents who seek to buy and sell properties. House prices emerge from this market process. Model Definition
According to McMahon et. all (2009) the main actors in the housing market (people, banks) and the components (houses, mortgages) are modelled as agents. People are either renters or owners of one or more houses. Mortgages are Adjustable Rate Mortgages (ARMs) based on interest rate. Each house is associated with zero or one mortgage that is owned by a bank. PeoplePeople have a fixed income that follows a uniform distribution within some range. They may relocate, which in turn requires to rent or own houses. This decision requires the evaluation of the financial situation with respect to the rent or ownership situation. House owner may evaluate the decision to buy an extra house for investment and rent it out to other people. House owner may decide to sell a house. HousesHouses can be rented or owned. Initial house prices follow a uniform distribution within some range. Houses are foreclosed, due to insufficient funds to pay mortgages or rents. MortgagesThe mortgage is owned by a bank, and is associated with a particular person and a house. Mortgage payments are adjusted to represent the notion of ARMs, possible with some time lag. BanksBanks maintain a balance sheet to keep track of their assets (mortgage payments) and liabilities (mortgage value of the houses owned by the bank). Model initializationHouses are created at a certain density (patchDensity) on the map. Each house is then assigned a price. A fraction of houses are assigned as rentals (rentalDensity) and some percentage is occupied (occupiedPerc) by people. Each person is then assigned an income and People are assigned to houses according to matching income  rental/ownership costs. The parameter patchDensity, rentalDensity, and occupiedPerc can be calibrated to empirical data.The model generates the following output: average house Price, average mortgage cost, number of owned vs. rented houses, banks balance sheet, and percentage of bankrupt people, and average location of houses. First ResultsWith the Axtell & Epstein (1994) model, we can observe patterns of development out of the individual interactions of agents qualitatively and try to test them with the empirical data. Using the actual interest rate scenario 19932009 a big drop in the banks balance sheets can be observed, that corresponds to the decrease in the house prices and the increase in the average mortgage rates. A policy of controlling exogenously for the interest rates leads to the emergence of bubbles and foreclosures. Things to doAs first results have shown, an exogenous control for the interest rates was one of the factors that led to the housing crisis. However, this is considered to be a necessity, but not a sufficiency condition: Future models should link the subprime crisis to the housing market. The model can be further enhanced by including the supply side of the market represented by the construction companies, in order to refine the endogenous emergence of the house prices. References:

Agentbased Computational Economics
Empirical Techniques to Probe Social Network Structure
Social networks are networks in which vertices represent people or groups of people and edges are social interaction among them, e.g. friendship. In sociological terms vertices are actors and edges are ties. The study of social networks goes back to 19th century psychiatrist Jacob Moreno who became interested in the dynamics of interactions within groups of people. Moreno [1934] called his diagrams sociograms which later become known as social networks. In his study on schoolchildren he used triangles and circles as vertices to represent boys and girls respectively. A friendship relationship is indicated by an edge connecting two vertices. The diagram reveals that there are many friendships between two boys and two girls, but few between a boy and a girl. Once drawn the diagram, it was easy to see. This is what persuaded social scientists that there was merit in Moreno's methods. Depending on the question one is interested in answering there are many different ways to define an edge in such an network. Edges may represent friendship, professional relationships, exchange of commodities, communication patterns, romantic or sexual relationships, or many other types of connections between individuals. The techniques to probe different types of interaction may involve direct questioning (e.g. interviews) [Rea97, Rap61], direct observation of experimental subjects, the use of archival records (e.g. the "southern woman study" [Davis41]), egocentered data analysis [Burt84, Bern89], affiliation analysis [Davis41, Gal85], smallworld experiments [Mil67, Trav69], snowball sampling [Erick78, Frank79, Thom00], contact tracing and random walk sampling [Klov89, Thom00]. This techniques have been applied to problems like friendship and acquaintance patterns on different scales of the population, e.g. students, professionals, bord of directors, collaboration of scientists, movie actors, musicians, sexual contact networks, dating patterns, covert and criminal networks such as drug users or terrorists, historical networks, online communities and others. A classic problem of social network analysis is to discover clustering. In the reminder of this article we will focus on different empirical methods used to measure social networks. InterviewsAsking people questions is the most common way to accumulate data about social networks. This can be done in the form of direct interviews, by questionaries or a combination of both  each with advantages and disadvantages with respect to the quality of data. A good introduction to social survey design and implementation is Rea and Parker [1997]. Surveys typically employ a name generator  a mechanism to invite respondents to name other nodes in the network as well as their relationship to them in order to explore the network. In the study of Rapoport and Horvath a question to the schoolchildren was to name eight best friends within the school. There are some interesting points to notice about name generators. Nominating other vertices by ties is an asymmetric process. Person A may nominate person B as friend but there is no need that person B has to nominate person A as friend. Therefore it makes sense to represent this data as directed networks. Vertexes in directed networks have two types of degree: indegree  the number of individuals who identified the vertex as friend  and outdegree  the number of friends identified by the vertex. The second point concerns the limit of responses given to the respondents. In the study above the limit was to name up to eight friends. Such fixed choice studies limit the outdegree of the vertices. This cutoff may lead to the loss of information about the smallworld effect of the network, which is caused by a small number of highly connected vertices. However, the indegree is not affected by such cutoffs. Studies based on direct questions are not only laborious, inaccurate and costly. Most of all, the data contain uncontrolled biases. EgoCenteredFor determining network structure, sosiometric studies  such as in the previous section require a survey of all or nearly all of the population. A reconstruction of the complete networks of ties is not possible. Given the high costs to survey large networks, a study of personal networks or egocentered networks may be a feasible alternative. An egocentered network is a network about one individual (ego) and its surrounding immediate contacts (alters). A typical survey would be to sample the population at random and ask them to identify all those with whom they have a certain type of contact. Also, they are asked for information about characteristics of themselves and there alters. This type of of survey is useful in particular if we are interested in the degree of the network. A random sample of degrees can give a reasonable degree statistics. In case we also gather information about contacts between alters, we can also estimate clustering coefficients. If we have data on characteristics of egos and alters we can estimate assortative mixing. ObservationDirect observation over a period of time is an obvious method to construct social networks. This is a rather laborintensive method. It is restricted to small groups, primarily ones with facetoface interactions in public settings. It is the only viable experimental technique for social network studies in animals. Archival DataA highly reliable source of social network data is archival records. Affiliation NetworksAffiliation networks are special kind of social networks to focus on cluster discovering. SmallWorld ExperimentSnowball SamplingContact TracingRandom WalksReferencesThe best general introduction on network theory is the book from Mark Newman [2010]. There is an active research community mainly somehow affiliated with the Santa Fe Institute.

Information Theory
Entropy and InformationInformation is reduction in uncertainty and has nothing to do with knowledge. Imagine you are about to observe the outcome of a coin flip and the outcome of a die roll. Which event contains more information? After Abramson, the information contained in the outcome of an event E with probability P(E) is equal to log 1/P(E) bits of information. For the unit bits we use log base 2. The result of a fair coin flip we get (log 2 = 1 bit) and for the die roll (log 6 2.585 bits). EntropyNow imagine a zeromemory information source X. The source emits symbols from an alphabet {x_1, x_2, . . . , x_k} with probabilities {p_1, p_2, . . . , p_k}. The symbols emitted are statistically independent. What is the average amount of information in observing the output of the source X?Shannon formulated the most fundamental notion in information theory for a discrete random variable, taking values from $\mathcal{X}$. The entropy of X is
Alternative explanations of entropy can be the average amount of info provided per symbol, average amount of surprise when observing a symbol, uncertainty an observer has before seeing the symbol, avg # of bits needed to communicate each symbol.
Proposition
InterpretationH[X] measures:
“paleface” problem Description LengthH[X] = how concisely can we describe X? Imagine X as text message:
Known and finite number of possible messages (#X). I know what X is but won’t show it to you. You can guess it by asking yes/no (binary) questions First goal: ask as few questions as possible
New goal: minimize the mean number of questions
Theorem: H[X] is the minimum mean number of binary distinctions needed to describe X. (Units of H[X] are bits) Multiple VariablesJoint entropy of two variables X and Y: Entropy of joint distribution: This is the minimum mean length to describe both X and Y
Entropy and Ergodicity(Dynamical systems as information sources, longrun randomness) Relative Entropy and Statistics(The connection to statistics) ReferencesCover and Thomas (1991) is the best single book on information theory. TutorialsConferencesResearch Groups

Processes and their Causal Architecture
The behavior of complex systems is a result of their internal structure. This structure reflects how processes compute information. This is determined by answering three questions:
intrinsic unpredictability (deterministic chaos) emergence of structure (selforganization) Main Idea
Readings

Cellular Automata
A onedimensional (elementary) cellular automaton consist of a row of cells. Each cell can be in, say one of two states  state '0' and state '1'. At each step there is a rule to update the state of each cell, say based on that cell and its immediate left and right neighbors. The table below is a representation of the rule for this kind of cellular automata. It is a kind of lookup table. The top row contains all possible state combinations for a cell and its left and right neighbor at step n. The botom row specifies the state of that cell at the next step in each of these cases.
