Oleg S.

Top Rated Plus | 100% Job Success | 5.0 from 49 reviews With 15+ years as a Data Engineer and an M.S. in Math, I am delivering data pipelines, data warehouses, dashboards, and machine learninig for eCommerce, Marketing, and Finance. 𝙎𝙪𝙘𝙘𝙚𝙨𝙨 𝙎𝙩𝙤𝙧𝙞𝙚𝙨 1. Payments forecast with MAPE < 2%: Banks need to know their monthly cash inflow to plan new credits and attract investments. I have compiled historical payments from existing clients, organized by calendar month, loan date, and credit product. Used this info to iteratively forecast new payments from new credits. 2. Data Quality from 24% to 99%: fixed core table with account balances in a banking DWH. Analytics accounts in accounting equal the sum of the basic accounts. I made a list of clients with analytics and basic accounts that mismatch. Picked one basic account with the most zero balances across all clients. Statistically, it had the most chances of being wrong. Discussed it with the Admin of the source system, and after several find-and-fix iterations, he confessed there is a filter that does not export all the data into the DWH, enforced by security guys. 3. Credit scoring models development, reducing scoring costs by half: A Fintech company was looking for software to build an internal credit score. They were considering SAS, SPSS, and other software. I proposed to them to build a scoring model using the free R-project, and I was hired immediately. Have joined applications and credit histories data and built scoring models using logistic regression. After the AB test with underwriters, the model proved its efficiency and was deployed in production. We stopped paying for credit scores from 3d party provider reducing scoring costs by 2 times. 4. I was optimizing the legal collection process for a bank. The overall collection process starts with contacting the debtor and discussing the payment plan. If the debtor is unavailable or refuses to pay, the legal collection process starts. Lawyers fill in applications for the Magistrates’ court, paying some court fees. The court may decline the application if it reaches the court too late due to the statute of limitations or take the debtor’s side. Even a won case does not guarantee the debtor will pay. He might decide to file for bankruptcy. Given the high number of debts and fees involved in the court applications, legal collection is a costly process. Multiple outcomes make this process manually unpredictable. I was tasked with building a collection scoring model to filter those debts that are worth taking legal actions. I have pulled information on 100K debts that went through the legal collection process past 3 years. My target variable was the sum of payments after legal actions minus legal fees. I have tried many predictor variables, and the most informative were: • debt amount - the higher the debt, the higher the payment amount is expected • days past due - the higher the DPD, the higher the chances of hitting the statute of limitations • debotr’s region - courts' practices vary across the country, with some courts taking debtors’ side more often Once the model has been built, we applied it to the current debts in the legal process and found out that only 30% of debts are worth taking to court. The rest 70% could be sold to collection agencies. The legal collection scoring model helped to understand: • the factors that drive collection profits • automate manual decisions • increase collection profits • clear debts balances, selling them to professional collection agencies 𝙏𝙤𝙤𝙡𝙗𝙤𝙭 • ETL/Data Warehouse: SQL, dbt, Python, Redshift, BigQuery, SAS • Data sources: Shopify, GA4, Google Ads, Meta Ads • Analytics dashboards: Looker Studio, Looker, Power BI, Excel, Google Sheets • Finance math: ROAS, Gross margin, LTV, Cash Recovery • Data Science and ML: Credit Scoring and payments forecasting • Cloud tech: AWS, GCP, Docker • Technical writing: Business Proposal, Product Description, User's Guide, Installation Guide • Project management: Agile / Scrum Let's connect to discuss the details.