Predrag Miocinovic

Designed, prototyped, and implemented a back-end engine for asset mix optimization and distribution of retirement holdings. The engine is implemented with Python/SciPy optimization logic wrapped around the Rust kernel performing a multiyear asset balance and tax calculation projections.

Using REST API, the engine is integrated with the front end, providing real-time graphical updates to projected tax obligations and wealth transfer amounts to a user modeling a senior retirement plan.

Insurers manage assets funded by premiums to guarantee sufficient liquidity for future claim payouts and maximize return on assets invested. A US insurer ($100 billion in AUM) needed an automated way to visualize the outcomes of its asset reinvestment rules under what-if scenarios in a multi-year forecast, which they had analyzed mainly in Excel.

I formed and led a team to implement an asset reinvestment optimization algorithm in an income forecasting simulation (KRM). This required: 1) implementing an algorithm in C++ to forecast selling and buying of assets following the reinvestment rules and constraints and 2) defining new data structures in an SQL database to configure such rules and constraints and store intermediate results of algorithmic decisions. The results of asset forecasts were integrated into the client’s Power BI framework.

By automating the execution of asset reinvestment rules in existing income forecasts, the client accelerated its processing of what-if scenarios and simplified the configurations needed to drive income forecasts. They could test and analyze the effects of new and modified reinvestment rules more quickly and precisely to select the best-performing assets for their trading strategy.

Kamakura Risk Manager (KRM) is an integrated risk management engine for the financial industry. It's fully scalable and deployable from a laptop to a cloud. As a lead designer and developer for 16 years, I worked on all aspects of this product.

FOCUS AREAS
• Addressed technology debt by redesigning the modeling library to use object-oriented design to model financial instruments.
• Implemented financial modeling and valuation for vanilla and exotic financial products.
• Designed a system for high-performance multithreaded computation and input/output.
• Expanded capabilities of Monte Carlo-based simulations of macroeconomic risk factors and instrument valuations.
• Enhanced the custom scripting interface for user-defined econometric variables.

Kamakura Risk Manager provides a scripting interface that allows clients to extend many of its calculations with their own custom calculations, including the ability to replicate exact calculations used in other risk management solutions for cross-checking and validation.

A large Central European bank (managing €275 billion in assets) migrated from a legacy risk system to KRM, using this scripting interface to replicate interest rate forecasting logic used by its legacy solution for validation in KRM. The client’s scripts used recursive logic for rates over a 10+-year horizon. When placed in production, several hundred of these scripts took more than an hour to parse and prepare for calculation.

Using Visual Studio profiling tools, I optimized a C++ script parsing algorithm to use more efficient STL-based memory objects, significantly improving the parsing performance. With the new algorithm, original scripts took only seconds to be parsed and prepared for calculation.

The bank was able to complete several production runs in the same time it took the original code to prepare for a single run. This allowed for quicker calculation comparison and validation, which led to a faster UAT process and acceptance of the new system.

A pioneering international scientific project needed a way to visualize data collected by optical sensors embedded in huge volumes of glacial ice. I developed an animated 3D-event viewer that allowed for zooming, rotations, and full customization of the visual display. At the back-end, I implemented a raw data parser that converted flat files into an object-oriented data library to enable easy manipulation and a complete display of all data properties in the application.

Animations and static 3D views of the data helped to guide data analysis and provided easily understood visuals. This allowed the project to popularize and communicate its mission and results among scientific peers, funding agencies (resulting in a multimillion-dollar allocation in the Congressional budget), and the wider community interested in scientific discovery (https://www.nature.com/articles/nature.2013.13026).

After discovering LIBOR interest rate benchmark manipulation, markets moved to abandon LIBOR in favor of risk-free rates (RFRs), and some LIBOR rates ceased to be published. This is a problem for long-term contracts underwritten with LIBOR benchmarks that expected LIBOR rates to remain available.

To support these contracts, the ISDA published rules to calculate LIBOR-like rates from RFRs. The rules satisfy original LIBOR logic and availability of RFRs for future calculations, but they are complex because defining how to construct a replacement benchmark rate must account for global differences in business days and holidays.

An Australian bank with 900 billion AUD in AUM and long-term LIBOR contracts required the ISDA rules for the LIBOR replacement rate when forecasting its long-term balance sheet positions. I implemented a C++-based algorithm that executes the rules, enables efficient recalculation of rates under simulated market shocks to RFRs, and allows for currency-specific rate differences. Despite more complex logic, simulations performed with the new algorithm have minimal speed degradation compared to original LIBOR-rate-based simulations. The client uses it in nightly production runs without workflow modifications.

Banks are required to hold sufficient liquidity and funding profiles to survive financial stress by meeting liquidity coverage and net stable funding ratio (LCR/NSFR) thresholds. To calculate these ratios, banks categorize assets based on 40+ criteria and perform specific aggregations and calculations.

In forecasts, banks must account for reinvesting maturing assets by simulating and categorizing new records on their balance sheets to include in LCR/NSFR values. Large banks have hundreds of millions of simulated asset records, so it’s untenable to categorize them across 40+ dimensions via legacy SQL-based workflows.

A major Southeast Asian bank ($90+ billion in AUM) had process bottlenecks and configuration limitations, thus calculating only current LCR/NSFR values. I used C++ algorithms to replace the SQL-based workflow for these calculations. By keeping the existing SQL categorization setup, the client didn’t have to modify its inputs and seamlessly used the faster routines. Thus, millions of simulated records were instantaneously categorized and aggregated, allowing the client to efficiently calculate and predict both current and future LCR/NSFR values and adjust funding and asset selection models to meet mandated limits.

A US agricultural lender ($14+ billion in loan volume) had to upgrade its expected-credit-loss modeling to comply with the current expected credit loss (CECL) methodology. In simulations, the lender had to calculate expected losses at the record level to match and be attributed to specific loans on its balance sheet.

Agricultural loans are atypical in structure and the variability and lumpiness in their payoffs. This lender has detailed models for future loan payoffs and default probabilities. CECL requires loan cash flows to be adjusted and expected losses calculated based on default probabilities in a very specific way that isn’t compatible with time-dependent forecasting methods.

To cover the incongruence between the regulatory requirement and modeling methodology, I implemented a new algorithm in the KRM net income simulation, which calculates CECL-adjusted cash flows and expected losses for any type and future timing of loan payoffs without time-consuming recalculation of loan amortization schedules. This allowed the client to use a single production run, greatly reducing the cost of server time and complexity of workflow to forecast both future income based on its models and CECL-based provisions for annual filings.