Laboratory Projects
Reinforcement Learning for Delta-Hedging in Illiquid Markets
This project considers the problem of delta hedging in illiquid markets, where transaction costs, limited market depth, and the market impact of large transactions play a significant role. Classical methods based on the Black–Scholes model assume continuous trading and infinite liquidity, which leads to significant distortions in practical application. To overcome these limitations, an approach based on reinforcement learning using the Bellman risk aversive operator is proposed. The environment used is an agent-based exchange simulator with support for trading the underlying asset and options, simulating the microstructure of the market and the dynamics of the market depth. A DeepLOB convolutional encoder is used to extract the features of the market depth, which makes it possible to take into account the hidden characteristics of liquidity. Numerical experiments show that the proposed method generates a distribution of the implemented PnL centred near zero and with less heavy tails compared to the classical Black–Scholes delta hedger. In addition, the risk-aversivity parameter λ enables management of the compromise between average profitability and tail risk control. The results demonstrate the effectiveness of the approach and its applicability for building sustainable hedging strategies in an illiquid market.
The Dynamics of Toxic Order Flow: Understanding Adverse Selection in Financial Markets
This project examines the impact of high-frequency trading (HFT) on the toxicity of the flow of orders and the dynamics of orders in financial markets. The analysis is based on MOEX exchange data on stocks with different levels of intraday trading volume—from large companies (blue chips) to medium and small ones in terms of intraday trading volume. Using the VPIN (volume-synchronised probability of informed trading) indicator as a measure of order flow toxicity, the study reveals the relationship between HFT activity and changes in the dynamics of liquidity supply and demand. The results show that VPIN is an effective tool for predicting changes in the supply and demand of liquidity from high-frequency traders. In addition, the impact of HFT on stock price volatility varies depending on the level of order flow toxicity and trading volume.
Image Generation Based on Text Descriptions while Preserving the Identity of the Subject
In recent years, diffusion models have proven to be the most effective approaches to image generation based on text descriptions. However, one of the key problems remains the preservation of the identity of a particular subject when transferring it to new visual contexts. This research is aimed at developing and analysing methods to integrate the representation of the subject into the architecture of diffusion models. Special attention is paid to balancing the preservation of the unique characteristics of the subject and the variability of the images created. An important aspect is also the choice of metrics for quality assessment: the structural similarity of images, perception of the subject's identity, and matching text queries. The solution to this problem has high practical value for the e-commerce industry, where personalised content generation can be used to automatically create advertising materials, facilitate virtual product fitting, generate creative assets for marketing campaigns, and improve the quality of the user experience.
In-Context Generation via a Single Image
This project is devoted to the problem of retraining during the refinement of individual modules of diffusion models on limited data, especially in the adaptation scenario for a single reference image (single-image customisation). A method is being developed that dynamically adjusts low-rank updates of the model weights depending on the current diffusion step. Projects on changing the effect on the model depending on the diffusion step are currently underway. Different patterns of behaviour are revealed: in some cases the model is more capable of retraining, and in other cases it is almost impossible. As part of the research, several options will be considered for sampling steps in the diffusion process and changing the effect of the loss function on the final result depending on the step.
Study of Price Manipulation Schemes in Cryptocurrency Markets. Analysis of Insider Buying Volume and Post-Event Price Discovery
In this project, we study the topic of detecting pumps and dumps of cryptocurrencies and build a classification model that can mark tickers susceptible to a pump up to one hour before it begins. We apply a new cross-sectional approach to problem formulation, as well as cross-sectional normalisation of features. We trained several boosting models and compared the results with other similar studies. We also offer a trading strategy that invests in potentially pump-prone assets based on the logs of our models.
Time Series Analysis under the Influence of Correlated Noise
The study is devoted to the analysis of time series subject to correlated (colour) noise and the assessment of how the spectral structure of noise affects the dynamics, predictability, and detection of regime shifts. SDE models (synthetic and based on real data) are built with added colour noise and white noise as a benchmark. The stability of transition points (PELT) under a varying noise spectrum is investigated. For forecasting, we use both linear basic models and nonlinear ones (LSTM, CatBoost), comparing the benefits of a realistic correlation structure of noise against white noise. Practical cases include asset returns, interest rates (Vasicek, CIR), and telemetry of systems where low- and mid-frequency correlations are critical. The results make it possible to improve early warning of regime change and increase the robustness of forecasts in financial and economic tasks.
Algorithmic Pseudo-Collusion in Continuous Double Auctions
In the study [Pastushkov, Boulatov, 2025], a result was obtained of the possibility of strategically underestimating the liquidity provided by learning agents in the financial market with a single price auction. This study develops a program to study algorithmic pseudo-collusion in financial markets. The double continuous auction is the most widely used mechanism in global financial markets, a fact that determines the relevance of this study. The paper will explore the possibility of algorithmic pseudo-collusion in markets with this type of market mechanism. The research also involves a quantitative comparison of the results.
Distributed Reinforcement Learning of a Competitive Agent with a Customisable Difficulty Level and Behaviour Style
In competitive online multiplayer arenas, there are no agents based on deep learning and reinforcement learning, in particular, those with the ability to customise the difficulty level or style of play. The goal of the project is to create such a tool. The resulting experience can be extrapolated to other systems that simulate human behaviour, as well as to training methods for similar systems or systems that require behavioural variability.
Investigation of the Possibility of Identifying Toxic Orders inside an Order Book
This research is devoted to the application of order flow analysis methods and market microstructure. The work uses proprietary high-frequency data with microsecond accuracy. Classical toxicity metrics such as Order-to-Trade Ratio, Order Flow Imbalance, Kyle's Lambda, VPIN, and others are used for quantitative assessment. In addition, machine learning methods (such as XGBoost) are used to predict market effects and unsupervised clustering algorithms are employed to identify different trading flow modes.
Analysis of the Behaviour of Financial Market Participants using Agent-Based Modelling
This study analyses the behaviour of financial market participants as complex adaptive systems with high dynamics. The aim is to study agents' reactions to shocks and price formation mechanisms. For this purpose, a 2D ABM simulator is being developed that simulates trading of multiple assets on multiple exchanges simultaneously. This makes it possible to explore inter-exchange effects: how an event on one site affects the trading of the same asset on others. A series of experiments is planned to identify factors of market stability and the adaptability of agents' strategies. At the next stage, reinforcement learning methods are integrated so that agents can independently learn and optimise trading solutions, which will reveal the mechanisms of participants' adaptation to changing conditions. The results are expected to deepen the understanding of market dynamics and find applications in financial engineering, strategy development and evaluation, and risk management.
Students of the HSE Faculty of Computer Science and other faculties are invited to participate in research projects, including course projects, bachelor's and master's research if the research area directly or indirectly coincides with the research areas of this laboratory. The top students will receive an offer for the position of research assistant.
If you have any questions about participation in the scientific activities of the Laboratory of Artificial Intelligence in Mathematical Finance, please contact the Head of the Laboratory—Peter Lukianchenko.