Data-Driven Financial Conduct Regulation: the FCA’s remit, datasets and research, and opportunities for collaboration Dr Stefan Hunt Head of Behavioural Economics and Data Science Big Data Analytics for Financial Services, UCL 7th January 2016 1 Remit of the FCA We regulate most of the UK financial markets. Retail: - Savings and investments - Consumer credit - Mortgages - Insurance - … and wholesale: - Investment banking - Fund management - … 2 Number correct as at 6 January 2016. Does not include consumer credit firms with interim permissions. “Other” firms are mainly consumer credit Objectives and powers Strategic objective Operational objectives Ensure that financial markets function well Market integrity Consumer Protection Promoting effective competition The FCA intervenes in markets through: • Authorising firms and people to operate • Policy-making: creating laws • Supervision: check compliance • Enforcement: prosecution and punishment …increasingly using competition analysis Key FCA data sets Wholesale: 1. Financial transactions / Zen 2. EMIR (interest rates, OTC derivatives) 3. AIFMD (hedge funds) Retail: 4. Payday lending 5. Credit card statements (~ all statements for last five years) 6. Credit bureau files 7. Personal current account micro data 8. Data from large field experiments (e.g. savings, insurance), matched with surveys 9. Product sales data (retail products, mortgages good quality) Firms and employees: 10.Firms’ regulatory submissions, consumer complaints etc. 11.Employees’ authorisations and records The data ecosystem Firm regular & ad-hoc submissions Supervision, Enforcement etc. Complaints & supervisory data Data Audit and ingest Elastic high-performance cloud storage Machine learning & statistical models Credit bureaus Surveys ONS Other Social media Visualisation Payday lending price cap Parliament created duty to impose cap on “high-cost shortterm credit”. Structure and level decided by FCA Questions: 1 What happens to firms and firms’ lending decisions? 2 What options are there for consumers without access to loans? Are they better or worse off? Data • Requested data using formal legal powers • Data on payday loans in 2012-3: top 37 lenders, ~99% market • For 11 lenders, ~90% market, all applications, denied and accepted, including lender credit score and revenues and costs • Match applicants across firms and to credit bureau files using unique identifier. 6 years of data including loan applications, holding and balances, credit events, defaults and credit bureau credit scores • Dataset of vast majority of first-time loan applications, ~1.9million applicants (observe 4.6 million people, ~10% of adult population) 7 Recreating lending decisions: credit scores ‘Good’ credit score 45o – credit score has no explanatory power ROC = Receiver Operating Characteristic 8 Recreating lending decisions: customer level profitability Migration matrix: 𝑚𝜏,1,1 ⋮ 𝐌𝜏 ≝ 𝑚𝜏,𝐵,1 Decision rule: 𝒂𝝉,𝟏 𝐀𝝉 ≝ ⋮ 𝟎 Loan profitability: ⋯ ⋱ ⋯ ⋯ ⋱ ⋯ 𝑚𝜏,1,𝐵 ⋮ 𝑚𝜏,𝐵,𝐵 𝟎 ⋮ 𝒂𝝉,𝑩 𝐌0 for Q0 (migrations to Q1), 𝐌1 for Q1 (migrations to Q2) and 𝐌𝟐+ for subsequent Qs (to following Q) ∗ 𝐌𝟐+ ≝ 𝐌𝟐+ 𝐀𝟐+ ∗ 𝐌𝟏 ≝ 𝐌𝟏 𝐀𝟐+ 𝐌𝟎∗ ≝ 𝐌𝟎 𝐀𝟏 𝝅𝝉,𝟏 ∀ 𝚷𝝉 ≝ ⋮ 𝝉∈ 𝟎,𝟏,𝟐+ 𝝅𝝉,𝑩 ∗ ∗ 𝟐 Customer profitability: 𝜽𝟐+,𝒃 ∶= 𝟏𝑻𝒃 𝚷𝟐+ + 𝒅𝒇𝟏𝑻𝒃 𝐌𝟐+ 𝚷𝟐+ + 𝒅𝒇𝟐 𝟏𝑻𝒃 𝐌𝟐+ 𝚷𝟐+ + ⋯ ∗ −1 𝚯𝟐+ = 𝐈 − 𝒅𝒇𝐌2+ 𝚷𝟐+ 9 𝚯𝟏 = 𝚷𝟏 + 𝒅𝒇𝐌𝟏∗ 𝚯𝟐+ 𝚯𝟎 = 𝚷𝟎 + 𝒅𝒇𝐌𝟎∗ 𝚯𝟏 Example: Impact on customer profitability Expected Customer Lifetime Profitability Before Cap After Cap Credit score 10 Use regression discontinuity design to estimate causal effect of payday loans 1st Stage: 2nd Stage: Probability of getting payday loan Probability of missing a nonpayday payment 80% 80% 40% 60% 0% 40% 250 500 750 5.9% causal impact of payday on missing payments 250 Internal Credit Score 11 500 750 Causal impact of payday loan use on consumers Change in likelihood of exceeding overdraft limit 95% confidence interval • Evidence suggests payday use worsens financial outcomes • Use behavioural models to assess welfare impacts Months relative to first loan application Next step: identify heterogeneous treatment effects, who is gaining and losing, using data science methods (Athey and Imbens, 2015) More practical examples of using research Retail: 1. Impact of annual summaries, mobile banking and SMS alerts in personal current accounts 2. Field experiments on information disclosures in savings and car and home insurance Wholesale: 3. Impact of high-frequency trading on institutional investors 13 Data Science Roadmap Machine-driven compliance Text Analytics Mis-selling or failure propensity Clustering Data Harmonisation Data collection & audit Feature Engineering Predictive Analytics Visualisation Proactive Regulation Summary • DATA: FCA collects rich transaction data + legal powers to gather more data • METHODS: Undertaken rigorous, ground-breaking empirical research to inform policies. Starting to use range of data science methods • PEOPLE: Empirical economists + data scientists • OPEN: Open to new ideas for research + collaboration. Regularly work with world-leading academics + aim to publish in top journals • ACCESSIBLE: Creating high-specification secure cloud environment facilitating off-site access 15 • REAL-WORLD RELEVANT: Research has to be immediately usable to inform policymakers > print(‘Thank you’)