Apr 2015 – Jul 2017 · 2 yrs 3 mos · Greater Chicago Area

• Developed and applied machine learning techniques for volatility forecast and also for studying behavior of variable annuities policy holders.
• Fund Services
• Developed models for forecasting stock volatility using 5-minute tick data from S&P500 as well as other available information. Worked with GARCH-type models using both maximum likelihood and Bayesian inference. Used Deep Learning for volatility modeling.
• Created Milliman MMRS performance predictive model based on market conditions.
• Implemented probability of backtest overfitting CSCV algorithm that Milliman uses for backtests.
• Variable Annuity Policy Holder's Behavior
• Performed industry-wide predictive modeling study of the variable annuities policy holder's behavior. The model predicts a probability that a person would lapse from policy based on the customer's behavior, personal characteristics and third party data (credit, health, marketing etc.)
• Designed clustering policyholders into groups using a custom KMeans algorithm.
• Implemented interactive model visualization using Power BI and RShiny.
• Developed a solution to use Apache Spark and H2O on the Amazon Web Services cluster.

Mar 2014 – Mar 2015 · 1 yr · Greater Chicago Area

• Mainly my work has been centered around Allstate's Big Data, its processing using Hadoop, model building, development of frameworks for analysis and usage of the models in production.
• Participated in the Kaggle's Allstate Purchase Prediction Challenge https://www.kaggle.com/c/allstate-purchase-prediction-challenge. Interviewed the winners and wrote a final report.
• Hadoop's MapReduce Framework . Used Python and R for Mappers/Reducers. Developed a
• MapReduce framework, where each job is fully controlled by configuration files. The output directory for any job contains metadata with various job information. Wrote a custom Hadoop partitioner and multiple output classes in Java for the MapReduce framework.
• Premium Estimation. Developed a process for a quick estimation of Allstate policy cost when changing several parameters without making time-expensive full premium calculation calls.
• As a result, we obtained a fairly fast way to estimate premiums for any coverage countrywide within few percent error.
• Model Building and Data Preparation. Built several predictive models such as customer's final purchase coverage, prior insurance policy cost, estimation of repair costs.
• I used various machine learning techniques, but mainly Gradient Boosting Machines(GBM).
• Parallel GBMs Research. As part of the research time I studied various implementations of
• parallel GBM algorithm, such as R GBM, H2O, SkyTree and XGBoost. I compared their
• results in terms of speed, performance and studying their advantages and disadvantages.
• Languages: Python, R, Bash shell, XSLT, Java
• Technologies: Hadoop, MapReduce
• Source Control: Git

Dec 2009 – Mar 2014 · 4 yrs 3 mos · Batavia, IL

• Project: NOvA experiment
• Tasks and Accomplishments:
• 1. Data Acquisition System(DAQ)
• Developed a Monte-Carlo simulation of the DAQ system
• Data formats model. Researched methods of C++ function delegates for data evolution model, improving performance of the initial implementation by a factor of 2
• Multi-threaded data taking and processing application with GUI
• 2. DAQ Cluster System Administration
• Setup and maintained a cluster of ~200 computers, working on Linux for parallel data processing in real time
• Developed a file backup system
• 3. Monte-Carlo Simulations
• Developed an application for MC simulation of particle propagation through matter
• Developed debugging tools for the MC simulations with a visualization GUI
• MC validations
• 4. Monte-Carlo Production
• In charge of MC production of large data sets (several TB) for data analysis
• Developed a Perl-based package to produce events in the detector
• 5. Data Analysis
• Involved in solving a problem of high-frequency electronic noise observed in the detector. Used FFT algorithm to obtain the noise frequency characteristics
• Designed and developed a Kalman Filter algorithm with a custom extension for 3D track reconstruction of particles undergoing a stochastic scattering in the propagation. Calibration parameters are extracted from a PosqreSQL database. http://goo.gl/neR12g
• Used multi-variate analysis techniques, such as Boosted Decision Trees and Neural Network for pattern recognition
• Applied linear regression and KNN for energy estimation
• Developed a set of selection cuts to identify problematic backgrounds caused by electronics failures
• Developed selection criteria for a neutrino search
• Involved in the implementation of a clustering algorithm
• A developer of the ROOT-based offline analysis framework
• I also supervised six graduate students.

Aug 2004 – Dec 2009 · 5 yrs 4 mos · Tuscaloosa, Alabama Area

• Project: MiniBooNE experiment
• Tasks and Accomplishments:
• Built a predictive optical model for the expected charge and time registered by photon detectors from observed event pattern.
• Designed and developed a reconstruction algorithm, based on a minimization of time and charge negative log-likelihood(NLL) for Neutral Current elastic events. The optimization is done by using MINUIT.
• Used pattern recognition techniques to identify signal events.
• Developed Boosted Decision Tree algorithm for classification.
• Performed various validations of the reconstruction and the data analysis techniques, which included building simple custom Monte-Carlo simulations of the events and toy models.
• Investigated various regularization techniques for the unfolding problem. Unfolding is closely related to regression, but used in case of grouped observations.
• Languages: C++, R, Bash shell, FORTRAN, Perl
• Technologies: R, ROOT (including multivariate analysis extension), Qt, C++ Thread, C++ Boost, SQL, CONDOR, GRID, XML
• Source Control: SVN, CVS
• Databases: MySQL, PosqreSQL
• OS: UNIX, Linux, MacOS
• IDE: Vim, RStudio, Netbeans, Eclipse