Data Analytics in Seoul, South Korea

Financial Services / Digital Banking / Data Analytics / Seoul

A Korean internet-only bank (18.2M customers — approximately 35 percent of the Korean adult population, KRW 24T in deposits, KRW 18T in loans, operating since 2017, growing from zero to Korea's fourth-largest bank by customer count in 6 years, with 1,200 employees and no physical branches) needed a customer intelligence platform to transition from rapid growth to profitable growth. The bank had acquired customers at extraordinary speed through: zero-fee accounts, higher savings rates, seamless KakaoTalk integration, and a UX that made traditional banking feel archaic. But growth-stage metrics were obscuring fundamental questions about customer profitability and long-term viability. Core challenges: (1) Customer profitability blindness — the bank knew total revenue and cost at the portfolio level but couldn't calculate individual customer profitability. With 18.2M customers and no branches, every customer interaction was digital — but the bank had 14 different internal systems tracking different aspects of the customer relationship (core banking, card processing, loan origination, investment platform, insurance marketplace, payment processing). Estimated that 40-45 percent of customers were unprofitable — using the bank as a secondary account for specific features (high-rate savings, zero-fee transfers) without generating sufficient revenue to cover acquisition cost and ongoing servicing. The bank needed to identify which customers were valuable, which were becoming valuable, and which would never be profitable — without the blunt instrument of imposing fees (which would contradict the bank's brand promise and drive churn). (2) Credit scoring for thin-file customers — the bank had deliberately targeted underserved segments: young adults (18-29, representing 38 percent of customers), gig economy workers, and consumers with limited credit history. Traditional Korean credit scoring (NICE, KCB — Korea's credit bureaus) had limited data on these segments, leading to either excessive rejection rates (losing potential good customers) or insufficient risk differentiation (approving customers who defaulted at rates 2-3x the portfolio average). The bank needed alternative data credit models — using: transaction behaviour, spending patterns, income stability indicators, and social/demographic signals to credit-score customers that traditional bureaus couldn't assess. MyData: Korea's MyData initiative providing additional data (with customer consent) from other financial institutions — but integrating MyData into credit models required analytics infrastructure that the bank hadn't yet built. (3) Product engagement fragmentation — the bank offered 12 products (savings, checking, loans, credit cards, investments, insurance, payments) — but average customer engagement was 1.8 products. Compared to traditional Korean banks where customers averaged 3.2 products, the internet bank was underperforming on cross-sell — partly because the product recommendation engine was basic (rule-based) and partly because the bank didn't understand which products each customer was likely to adopt. Low product penetration meant low switching costs — a customer with only a savings account could leave instantly (and 22 percent annual churn rate suggested many did). (4) Real-time personalisation at scale — 18.2M customers generating 45M daily app interactions (logins, balance checks, transfers, payments, product browsing). The bank wanted to personalise every interaction — right offer, right time, right channel — but its existing analytics ran in batch (overnight processing), meaning personalisation was always 24 hours behind customer behaviour. Korean consumers expected real-time responsiveness (ppalli-ppalli culture) — a customer browsing loan products at 10 AM should receive a personalised loan offer by 10:05 AM, not in tomorrow's push notification. KakaoTalk integration: the bank's primary communication channel being KakaoTalk (not email or SMS) — personalisation needed to work within KakaoTalk's message formats, API constraints, and channel-appropriate tone. (5) FSC regulatory compliance — the FSC requiring internet-only banks to demonstrate: responsible lending practices (consumer protection), adequate risk management (capital adequacy), and fair treatment (no algorithmic discrimination). The bank's use of alternative data for credit scoring needed FSC approval — the regulator ensuring that AI/ML credit models didn't discriminate against protected classes. MyData compliance required specific data handling, consent management, and portability capabilities that created additional analytics governance requirements.

We built a customer intelligence platform over 12 weeks — reflecting Korean speed expectations — providing individual customer profitability, alternative credit scoring, real-time personalisation, and regulatory-compliant analytics. (1) Customer profitability engine: individual-level economics across all products. Activity-based costing: calculating actual cost-to-serve for each of 18.2M customers across all 14 internal systems — digital interaction costs (app usage, API calls, customer service chatbot, human support escalation), financial costs (cost of funds for deposits, expected loss for loans, interchange for cards), and allocated costs (technology infrastructure, compliance, marketing acquisition cost amortised over expected customer lifetime). Customer P&L: monthly profitability statement for every customer, updated within 48 hours of month-end. Revealing: 42 percent of customers were indeed unprofitable (versus the 40-45 percent estimate), but the distribution was more nuanced than expected — 15 percent of "unprofitable" customers were in their first 12 months and trending toward profitability, while 12 percent were chronic money-losers who would never become profitable at current product usage. Lifetime value prediction: machine learning models predicting 3-year customer lifetime value based on: current product usage, engagement trajectory, demographic profile, and comparable customer cohort analysis. Identifying "future VIPs" — customers currently unprofitable but with behavioural patterns matching those who became highly profitable — for proactive relationship investment. Strategic segmentation: customers classified into actionable segments — "grow" (valuable and expanding), "nurture" (potentially valuable, needs product deepening), "maintain" (modestly profitable, stable), and "manage" (unprofitable, consider cost reduction). Each segment receiving different treatment strategies. (2) Alternative credit scoring: ML models for thin-file customers. Transaction-based scoring: credit models using 6 months of transaction data (with customer consent) — income stability (regularity and consistency of deposits), spending discipline (ratio of discretionary to essential spending), financial management signals (maintaining positive balances, avoiding overdrafts), and peer comparison (behaviour relative to demographic cohort). MyData integration: integrating MyData feeds from customers' other financial institutions — seeing a complete picture of the customer's financial life, not just the slice visible through the internet bank. MyData providing: income from employer payments at other banks, existing loan commitments, insurance premiums, and investment holdings. Model validation: alternative credit models backtested against 3 years of loan performance data — demonstrating that transaction-based scoring reduced default prediction error by 31 percent compared to bureau-only scoring for thin-file customers. Fairness testing: comprehensive bias testing across protected attributes (age, gender, region, income level) — ensuring FSC compliance with anti-discrimination requirements. The models achieving comparable or better approval rates for historically underserved segments while maintaining risk discipline. Dynamic scoring: credit scores updated in real-time as new transaction data arrived — not a static score calculated at application time, but a continuously updated risk assessment. A customer whose financial behaviour deteriorated received proactive risk management (credit limit adjustment, payment plan offers) before they defaulted. (3) Real-time personalisation engine: 45M daily interactions personalised. Event-driven architecture: every customer interaction triggering real-time analytics — app login, balance check, transfer, payment, product page view — each generating an event processed within seconds. Customer context: at the moment of each interaction, the personalisation engine knowing: current product holdings, profitability segment, recent behaviour, credit eligibility, and propensity scores for each product. Offer decisioning: real-time decision engine selecting the optimal next action for each customer — "this customer checked their savings balance, has KRW 15M in deposits earning 2.5 percent, is eligible for a time deposit at 3.8 percent, and has a 67 percent propensity to accept — present time deposit offer within 5 seconds." KakaoTalk personalisation: personalised messages delivered through KakaoTalk using the bank's official channel — message content, timing, and frequency optimised per customer. Korean-language message generation ensuring appropriate formality (formal polite for older customers, casual polite for younger demographics) and message length optimised for KakaoTalk's UX. A/B testing: continuous experimentation — every personalisation decision an experiment, with results feeding back into the models. The platform running 200+ simultaneous experiments across different customer segments, channels, and offer types. (4) Product deepening analytics: cross-sell intelligence. Product affinity: analytics identifying which product combinations were most likely for each customer — "customers with savings accounts who receive regular salary deposits have 4.2x higher propensity to adopt a credit card than average." Journey mapping: understanding typical product adoption sequences — most customers followed predictable paths (savings then card then loan, or savings then investment then insurance), enabling proactive offers at the right point in the journey. Churn prediction: identifying customers at risk of churning (reducing balances, decreasing login frequency, removing the app from primary screen position) — with intervention recommendations: "this customer's engagement is declining — offer a deposit rate bonus of 0.3 percent for 3 months to maintain the relationship, estimated retention value KRW 840K." Win-back: for customers who had already reduced engagement, analytics-driven win-back campaigns — personalised offers based on the customer's historical behaviour and the product most likely to re-engage them. (5) FSC compliance and governance: regulatory-grade analytics. Model governance: complete model documentation meeting FSC requirements — model development methodology, training data description, feature importance analysis, bias testing results, and ongoing monitoring plan. Model risk management: all production models having documented model risk assessments, validation reports, and performance monitoring dashboards with automatic alerts when model performance degrades. Explainability: credit decisions explainable to customers (FSC consumer protection requirement) — "your loan application was declined because: insufficient income stability (contributing 45 percent to the decision), limited credit history (30 percent), and high existing commitments relative to income (25 percent)." Audit trail: every analytics-driven decision logged with: input data, model version, decision rationale, and outcome — enabling FSC examination of algorithmic decision-making. MyData compliance: consent management, data portability, and access controls meeting MyData regulatory requirements.

Results over 12 months. Customer profitability: individual customer P&L enabling targeted strategies — "manage" segment (12 percent of customers, chronic money-losers) addressed through: cost-to-serve reduction (migrating to chatbot-only support, reducing marketing spend), resulting in 4 percent of these customers becoming marginally profitable and 3 percent churning voluntarily (desirable churn). Portfolio-wide profitability improving from 0.8 percent ROA to 1.2 percent ROA — KRW 96B (approximately USD 72M) improvement in annual net income. Credit scoring: alternative credit models approving 34 percent more thin-file customers while maintaining the same default rate — expanding the creditworthy customer base by 1.2M customers. Default prediction improving 31 percent for thin-file segments — reducing credit losses by KRW 42B annually. MyData integration providing credit decisions for 2.8M customers who had previously been declined or offered unfavourable terms. Personalisation: real-time personalisation driving product adoption from 1.8 to 2.4 products per customer — a 33 percent improvement. KakaoTalk personalised offers achieving 3.2x higher conversion than generic messages. Revenue per customer increasing 28 percent through better product matching. Churn reduction: annual churn from 22 percent to 16 percent — retaining 1.1M customers who would have otherwise left, with estimated lifetime value of KRW 1.8T (approximately USD 1.35B). Early intervention preventing 68 percent of predicted churn when intervention occurred 30+ days before expected departure. Cross-sell: product deepening generating KRW 180B in additional annual revenue — primarily from: time deposit conversion (KRW 62B), credit card adoption (KRW 48B), personal loan origination (KRW 45B), and investment product adoption (KRW 25B). FSC compliance: full regulatory compliance achieved — FSC examination confirming: alternative credit models met fairness requirements, algorithmic decision-making was explainable, and MyData handling was compliant. Zero regulatory findings in 12 months. Financial impact: profitability improvement KRW 96B; credit loss reduction KRW 42B; cross-sell revenue KRW 180B; churn value preservation KRW 1.8T (lifetime, not annual). Annual quantifiable impact: KRW 318B (approximately USD 238M). Development: KRW 1.2B (approximately USD 900K). Annual platform: KRW 380M (approximately USD 285K). ROI: 265x first year.