Konstantins Starovoitovs

Work experience

• 2019 – 2020: Pricing Model Validation @ Deutsche Bank (Berlin)
• 2018 – 2018: Software Engineer @ chessable (UK)
• 2017 – 2018: Software Engineer @ affilicon (Cologne)
• 2014 – 2016: Software Engineer @ ProPerforma (Cologne)

Research

• U. Horst, D. Kreher, and K. Starovoitovs. "Second-Order Approximation of Limit Order Books in a Single-Scale Regime," 2023. URL: https://arxiv.org/abs/2308.00805.
• B. Hambly, D. Kreher, and K. Starovoitovs "Non-negative Martingale Solutions to the Stochastic Porous Medium Equation with Sticky Behavior." (forthcoming)

---

Personal projects

Outside of my academic studies, I engage in projects centered around machine learning and quantitative finance. The overview of the recent projects can be found below.

• CVaR-optimal allocation for liquidity providers with a chance constraint

  Mar 16, 2024

  In this article, we explore a DeFi algorithm solving the automated market maker’s (AMM) problem, providing an optimal allocation of the liquidity provider’s (LP) funds between liquidity pools with varying trading characteristics. We implemented a simulated AMM environment and used the generated trajectories in the gradient descent iteration to find the CVaR-optimal allocation of the funds between the liquidity pools, subject to certain profitability of the strategy, expressed as a chance constraint. The mechanics of the AMM requires careful computation of the returns, and the problem is complicated by the probabilistic nature of the constraint. We address both problems by carefully designing the loss functional, and make use of the automatic differentiation to obtain the optimum. Finally, in order to account for the market impact of the individual liquidity provider, we alternate the simulation and gradient descent iterations.

• Generative modelling of sea surface temperature with normalizing flows

  Feb 11, 2023

  In this post we look at the deep flow-based generative model of the evolution of ocean temperature, based on real data provided by Mercator Ocean. We aim to learn the distribution of sea surface temperature (SST) data, including multidimensional dependencies, and generate potential future values of SST with spatial dependency between stations, in particular to simulate extreme climate scenarios within the context of stress testing and, more broadly, climate risk management.

• Domain adaptation with an adversarial algorithm for blood cell classification

  Dec 24, 2022

  In this article, we’ll present a machine learning model for classification of blood cells in hematology. We’ll emphasize how this algorithm leverages image recognition technology and, importantly, utilizes domain adaptation techniques to effectively handle variations in images from different labs, resulting in more accurate and practical diagnostics.

• Deep solver for FBSDE with jumps

  Nov 17, 2022

  In this article we elaborate on the deep solver we came up with the colleagues from HU Berlin for the solution of the stochastic control problem including jumps, as part of the Helmholtz GPU Hackathon 2022. Generally, any stochastic control problem can be rewritten as a forward-backward SDE (FBSDE), presenting an alternative to the dynamic programming approach with Hamilton-Jacobi-Bellman equations. Such systems are widespread in mathematical finance, arising in pricing of contingent claims, risk management problems and calculations of value adjustments (xVA) to account for the counterparty risk. In this article, we extend the deep FBSDE solver by E, Han and Jentzen, by making a deep ansatz for the control process $R$ corresponding to jumps, by analogy with the deep ansatz by E et al. for the control process $Z$ corresponding to the diffusive part.

• Human action recognition with log-signatures

  Jun 10, 2021

  In this article we look at the signature-based algorithm for skeleton-based human action recognition implemented by our team of colleagues from Berlin and Oxford, as part of the ICCVW2021 MMVRAC competition. We implement the PT-Logsig-RNN model by Liao et al., which combines extraction of the log-signature with convolutional and recurrent modules to transform the spatio-temporal skeletal data.

• Deep order book prediction model

  Mar 21, 2021

  We build a deep model for prediction of market moves based on the recent order book history. The model is based on the DeepLOB paper and consists of the convolutional and recurrent elements. A sequence of convolutional layers enables automatic feature learning, while the recurrent module captures temporal dependence. Portfolio constructed based on the model predictions leads to positive long-term P&L on the testing dataset, modulo transaction fees.