New Estimates of Trade Mis-invoicing and Methods based on artificial intelligence 26 June 2018 Kathy Nicolaou-Manias Armand Desjardines Overview of the Presentation Presentation Overview 1 Introduction 1. Introduction & Background 2 The Present 3 The Future 2. The Present 3. The Future 4 The Journey 5 Conclusion 4. The Journey 5. Conclusion Introduction and Background Measuring a hidden phenomenon Illegal Legal Introduction and Background Despite considerable research defining IFFs is a challenge White Market/Economy: Formal Laws, Regulations & Authorized Enforcement Transparent Formal Economy Penalties govern Taxed, monitored, Observed Formal Agents included in GNP Reported Regulated Legal/Licit engaged in Recorded Production & First Exchange Informal Economy/Sector Grey Market: Not taxed, monitored, included in Unauthorized any GNP, Unreported Employment (e.g. Spaza Shops, Shebeens, Taxis Shadow ARS, Hawala, Barter, Cash, DIY, Informal Self-Employment, Payment in Kind, etc.) Unobserved Unofficial Legitimate Unreported Unrecorded Criminal Acts Unreported (e.g. ML/TF, Unofficial Illegitimate Economy Second Counterfeits, ID (e.g. Tax Gap) Theft, Hacking, Opaque Informal Economy 419/Phising) Black Market: (e.g. Contraband, Unrecorded Piracy, etc.) Economy Unauthorized (e.g. Income/ Illegal Economy (e.g. Drugs, Informal Agents Production Irregular Arms, Organized Crime, Accounts, Social Circumvent, Escape & Sanction Busting, etc.) Shadow Grant Corruption, Excluded from the Government Budget Deficits, Informal Institutional Systems Corruption, Bribery, TBML) Underground Illegal Tender of Laws, Rules, Practices, Nepotism, Illegal Rights, Patronage, etc.) Unobserved Unregulated Illegal/Illicit Regulations & Unreported Enforcement Unrecorded Penalties Source: South African Financial Intelligence Unit: IFF project, 2012 Second Introduction Introduction and Background Micro Meso Macro FIC’s IFF unit looked at IFFs from the top down and bottom up bottom and down top the from IFFs at looked unit IFF FIC’s The Present UNECA Paper to Estimating IFFs: Approach 1. Overview • An overview of the approach describing the high level approach to the problem of quantifying IFFs. 2. Data • A descrption of the data used in the development of the model and an analysis of how the pros and cons associated with the choice of data. 3. Methodology • A descrition of the structure and statistical techniques used in the model. 4. Results • A review of the model results, usually based on the estimated magnitude of IFFs generated by the model. The Present The Data Environment In assessing the data environment for IFFs, four levels of data were identified: Country Level System of National Accounts Data Aggregated data based on the System of National Accounts Country Level Official Statistics Data (SNA), which is the internationally agreed standard Typically aggregated and set of recommendations on reported by the statistical Commodity/Sector Level (Official) Data bureau of the country, or other how to compile measures of Data collected for specific entity such as the South African economic activity. categories of goods and Transaction Level Data Revenue Authority, services or specific industries or representing official statistics from specific government e.g. quarterly basis by the South for the country. It analyses Data collected at individual sources. However, this data is African Reserve Bank (SARB) sectoral data that is supplied to transactions level. This data is analysed and becomes an with all national accounts the South African Reserve Bank typically very large and contains analytical data set. information in order to compile the SNA for significant amounts of detail. the country. e.g. large scale government e.g. data collected from banks administrative datasets (or e.g. IMF data, government and other financial institutions, samples thereof); Country-by- trade data, UNCOMTRADE cross border data flows held by Country reports to the OECD; data, sectoral production (or the SARB, customs data, tax Country Reporting Standard statistics on returns data, etc. submitted to the OECD; data on profitability/income after tax) gold exports from South Africa etc. or an MNE dataset; Corporate entities tax return datasets; etc. The Present Evaluating and Scoring the Existing IFF Models The TRAC-I Scoring Framework • Ease to replicate • A scoring framework was developed with Timeliness • Data availability the following five core pillars: • Model structure • Timeliness; • Developing economy context • Channels of IFF Relevance • Relevance; • Types of IFFs • Accuracy; • Data accuracy • Coverage; and • Statistical Techniques Accuracy • Limited assumptions • Impact • Minimal sensitivity • These five pillars represent the key • Multi-country application attributes of a universal IFF indicator. For Coverage • Effort to extend model to other each pillar, a generic score ranging from 1 countries (lowest) to 5 (highest) is possible. • Generates useful policy info Impact • Generates powerful diagnostic information The Present The IFF Scoring Framework for Data The key challenge in estimating and measuring IFFs is the availability of compressive, detailed and accurate data. • Readily available Availability • Limited barriers to access A Data Quality Framework was also constructed, comprising: • Multiple years • Availability of required data; Quantity • Disaggregated • Quantity of required data; • Universality of data across countries; and • Available in all countries Universality • Granularity of data • Consistent across countries • Multiple layers Granularity • Sector/product specific The Present Overview of the Finding – Methodology and Data Scoring of IFF Methods The Future Introducing Machine Learning and Artificial Intelligence “Machine learning (ML) is a method of data analysis that automates analytical model building. It is a branch of Artificial Intelligence (AI) based on the idea that systems can learn from data, identify patterns and make decisions with minimal human intervention - SAS” • Machine Learning has introduced Deep Learning increasing the sophistication of ML and its applications. • Financial and Banking sectors use machine learning to fight fraud and money laundering. • Research suggests that the adoption of machine learning based intelligence could drastically improve a countries ability to identify, measure and prevent illicit financial flows. • Existing machine learning solutions in the tax evasion and money laundering space have been developed and implemented, however this is mainly in private sector institutions. The Future ML and AI: A Solution for Estimating and Curbing IFFs in the Future Traditional Models Machine Learning (and AI) Models • Rely on proven theory to detect • Uses computing power to probe the data for new and known patterns or distributions in potentially unknown patterns in data data • Data modelling has been restricted to • With AI technology and Machine Learning, it is the detection of theoretically known possible to sift through massive amounts of data in phenomena within a dataset. The aim search of the unknown. was to fit a model to the data. • Illicit transactions are designed to • Machine Learning provides a potential solution to this pass undetected as far as possible, challenge where known and unknown patterns can be making their detection difficult, time detected, including the slightest deviation from these consuming and case-specific patterns, and can do so at speeds unattainable by human beings or even traditional statistical models. • By training machine learning models to recognise known patterns and unknown patterns of illicit trade, there is huge potential for improved IFF detection, quantification and eventual reduction. • ML techniques can augment/upgrade existing IFF models The Future Machine Learning – Understanding the Elements and Appropriateness for IFFs Which branches of AI are most suitable for IFF modelling? Level 1 (Machine Learning) Which approach to machine learning is most suitable for IFF Level 2 modelling? Level 3 Which types of learning are most suitable for IFF modelling? Which applications of machine learning are most suitable for IFF Level 4 modelling? Which machine learning algorithms are most suitable for IFF Level 5 modelling? The Future Machine Learning – Understanding the Elements • The core feature of any Machine Learning model is the ability to learn. • This learning can take place explicitly, where a body of knowledge is taught to the machine, or implicitly, where the machine begins to learn independently of the explicit information. This forms the primary difference between supervised and unsupervised learning. Learning takes place at different stages of a models lifecycle, The Future Examples of Machine Learning Solutions in the IFF Space There are a few examples of ML solutions applied to IFFs as shown below. Bare in mind that these solutions are geared towards detection of IFFs and therefore operate in real time, detecting illicit transactions as they occur. These solutions can be used to identify illicit transactions as they occur or can be applied to historical data to measure the magnitude of IFFs Case Study 1 Case Study 2 Case Study 3 • Title: Credit card fraud • Title: Neural Network Analysis of • Title: Neural network models: forecasting model based on International Trade Foundations and applications to an clustering analysis and • Author: Isaac Wohl and Jim audit decision problem integrated support vector Kennedy • Author: Rebecca C. Wu machine • Year: 2018 • Year: 1997 •