Work in An Agent-Based Tax Evasion Model Calibrated Using Survey Data Progress Attila Szabó1,2, László Gulyás1,2, István János Tóth3 1AITIA International Inc, Eötvös University, Budapest 3Institute of Economics , Hungarian Academy of Sciences 2Loránd Shadow 2011, Münster, Germany Shadow 2011, Münster, Germany 1 Acknowledgements The partial support of the following grants of the Hungarian Government are gratefully acknowledged. KMOP-1.1.2-08/1-2008-0002 OTKA T62455 OTKA K48891 Shadow 2011, Münster, Germany 2 Outline 1. The TAXSIM Model 2. Earlier Results 3. Survey-Driven Calibration (work in progress) 4. Results 5. Conclusions and Future Work Shadow 2011, Münster, Germany 3 The TAXSIM Model (since 2007) -“All models are wrong; some models are useful.” Box Shadow 2011, Münster, Germany 4 Previous models • The focus is on the taxpayers' strategy: Publications: Balsa et. al., Bloomquist, Davis et. al., Korobow et. al., Mittone and Patelli • TAXSIM makes a utilitarian approach: misreporting is a cost optimization strategy Shadow 2011, Münster, Germany 5 Actors of TAXSIM Employee (taxpayer) Try to find a job according to its preference Employer (taxpayer) Try to find employees according to its preference Tax authority Audits actors using a certain probabilistic strategy Government ('society feedback') Provides a level of contentment (services & satisfaction) Market sector (a single sector economy) Buys cheapest products first Shadow 2011, Münster, Germany 6 Model Architecture Shadow 2011, Münster, Germany 7 TAXSIM Decision Making Information from social networks Audits, Governmental services Taxpayer strategy Decision Model of the market Shadow 2011, Münster, Germany 8 Refinement of the Environment The taxpayers make a decision on … ...what they actually can! Taxpayer strategy respects the viable options (the environment may override the strategy) Employers and employees different strategies for different goals different motivation factors (government agent, model of the market sector) model of the labour market (how employers and employees cooperate) Shadow 2011, Münster, Germany 9 TAXSIM is... • … a simulator that aggregates Law enforcement Motivation towards compliance Taxpayers' goals • … rather a family of models than a simple model Shadow 2011, Münster, Germany 10 Earlier Results -“The only thing new in this world is the history that you don't know” (Harry S. Truman) Shadow 2011, Münster, Germany 11 First Set of Results • Starting from a ‘pessimistic’ sector, we investigated three scenarios: • Improving governmental services (taxpayers more often override optimal decisions) • Voluntary shift to total legalization (one company, cca. 15% market share) • Preferential taxes for companies (can afford higher wages at legal employment type) 200 180 160 140 Employed 120 100 80 60 40 20 80 0 10 00 12 00 14 00 16 00 18 00 20 00 22 00 24 00 26 00 28 00 30 00 32 00 34 00 36 00 38 00 40 00 42 00 44 00 46 00 48 00 50 00 52 00 54 00 56 00 58 00 60 00 60 0 40 0 0 20 0 0 Employed Legal employments Mixed employments 200 190 180 170 160 150 140 130 120 110 100 90 80 70 60 50 40 30 20 10 0 0 500 1000 Employed Employed 200 190 180 170 160 150 140 130 120 110 100 90 80 70 60 50 40 30 20 10 0 100 Employed 200 300 400 Legal employments 500 600 700 2000 2500 3000 800 Tick Mixed employments 900 3500 4000 4500 5000 5500 6000 Tick Employed 0 1500 Hidden employments 1000 1100 Hidden employments 1200 Legal employments Mixed employments Hidden employments 200 190 180 170 160 150 140 130 120 110 100 90 80 70 60 50 40 30 20 10 0 0 500 Employed 1000 1500 2000 Legal employments 2500 3000 3500 4000 Tick Mixed employments 4500 5000 5500 Hidden employments 6000 Studied the effect of social network topologies • Two-dimensional grid spatial/geographical • ErdÅ‘s-Rényi random graph (used in original studies) small world • Watts-Strogatz network spatial/geographical small world clustered • Real topologies depends on the modeled sector (construction, telco, etc.) Compliance depends on social network topology • Using random graph or Watts-Strogatz network results converge to the same equilibrium, but W-S is faster • Using a two dimensional grid agents don’t reach the optimal solution Survey-Driven „Calibration” -”To find out what happens when you change something, it is necessary to change it” (anonymous) Shadow 2011, Münster, Germany 16 Motivation • The Institute of Economics, Hungarian Academy of Sciences carried out a survey among Hungarian tax payers • We wanted to see how much TAXSIM can be made compatible with empirical findings. • Also, we wanted to make the behavior of the Tax Authority more realistic (i.e., adaptive vs fully random) Shadow 2011, Münster, Germany 17 Selected Findings of the Survey • Survey data shows that Hungarian tax payers estimate the probability of discovered tax avoidance to be 64% • Real audit probability is 3% • 73% of the respondents claimed that would comply even if audit probability was 0% Shadow 2011, Münster, Germany 18 Adaptive Tax Authority • The Hungarian Tax Authority assigns tax payers into three categories for audits: • Not audited yet • Audited and complying • Audited, non-complying (Further, sector and income specific categories are not applicable) • In TAXSIM the adaptive TA selects • the neighbors (in the social net) of the audited, non-complying agents. Shadow 2011, Münster, Germany 19 Survey-Driven Expectations 1. Citizens overestimate audit probability 2. There is about 73% of agents who would fully comply (pay taxes) even with an audit probability of 0% 3. Adaptive audit strategy increases compliance 4. Atomized society would yield a larger shadow economy (i.e., the social network is important) Shadow 2011, Münster, Germany 20 Results -”The best time to plan an experiment is after you have done it.” R. A. Fisher Shadow 2011, Münster, Germany 21 Citizens Overestimate the Probability of Audits • Not observed • Agents respond to the acts of discovery by the Tax Authority (on them or on agents in their network). However, there is no adaptation act responding to succesful non-compliance. • Agents learn the audit probability almost perfectly. Shadow 2011, Münster, Germany 22 High Compliance in Face of No Audits • Not observed • Agents are cost optimizers. This is a strong motivation for non-compliance. • The two opposite motivators are audits and satisfaction with governmental services. • The latter was shown earlier to be enough to improve compliance • Remark: the expectation was never observed (but claimed by respondents) Shadow 2011, Münster, Germany 23 Adaptive Audit Strategy Improves Compliance • Opposite results: adaptive Tax Authority yields higher level of non-compliance. • The simulated society turns illegal much faster. • Agents continue to see non-compliance as a cost optimization tool. • Audits and fines are the cost of this tool. • Adaptivity increases the efficiency of the Tax Authority, but this only increases the costs of „production”. Those who are not „illegal enough” cannot afford this and go bankrupt. Shadow 2011, Münster, Germany 24 Adaptive Audit Strategy Improves Compliance Shadow 2011, Münster, Germany 25 Adaptive Audit Strategy Improves Compliance Shadow 2011, Münster, Germany 26 Social Networks Matter • Confirmed • Less information, more non-compliance (remember, cost optimization is the main driver!) Shadow 2011, Münster, Germany 27 Conclusions and Future Work -”Perfection is not possible; it's always approximation” (anonymous) Shadow 2011, Münster, Germany 28 Conclusions and Future Work • Only 1 in 4 expectations confirmed • Overestimation could perhaps be achieved if we let the agents develop „false confidence” • Clearly, cost optimization is a very strong driver in this model • Yet, we intend to find the „turning point”: i.e., the level of governmental services that would make agents comply even without audits • The adaptive strategy made the Tax Authority more efficient, but fines were kept fixed • Thus, it was possible to „price in” the fines. Further parameter studies are warranted. Shadow 2011, Münster, Germany 29 Thank You! http://taxsim.mass.aitia.ai aszabo@aitia.ai lgulyas@aitia.ai Shadow 2011, Münster, Germany 30