1.
Contemporary Financial Regimes Modelling and Identification
- Financial markets experience structural breaks driven by macroeconomic shocks, monetary policy shifts, liquidity crises. This research theme focuses on moving beyond traditional models from hidden state-space literature by employing advanced non-linear frameworks to identify latent, unobserved market states in real time. The primary objective is to dynamically map parameters like volatility, drift, and cross-asset correlations to prevailing market regimes, mitigating model degradation during sudden structural transitions.
2. Causal Financial Machine Learning
-
Standard machine learning models optimize predictive accuracy in regression and classification tasks via historical statistical associations, making them highly vulnerable to spurious correlations. In asset management, the search for causal networks ensures that portfolio decisions rely on invariant features that remain robust under counterfactual interventions or shifts in macroeconomic policy.
3. Synthetic Financial Data Generation (focus on Synthetic Codependence Metrics Generation)
-
Training robust asset management algorithms often suffers from data scarcity, particularly regarding tail-risk events and changing market structures. This research theme investigates deep generative architectures—such as Generative Adversarial Networks and Normalizing Flows—tailored for financial time series. Special emphasis is placed on preserving joint distributions and synthetic codependence metrics generation (e.g., dynamic correlation matrices, tail-dependence coefficients, copula structures), ensuring that artificial data faithfully reproduces cross-asset co-movements and contagion effects without leaking proprietary information.
4. Deep Reinforcement Learning for Dynamic Asset Allocation
-
Dynamic portfolio optimization is formulated as a sequential decision-making problem under uncertainty, naturally suited for Reinforcement Learning (RL). This research theme addresses state-space representation, reward function design and sample efficiency. By leveraging Deep Q-Networks (DQN), Proximal Policy Optimization (PPO), and model-based RL, agents learn adaptive trading policies that directly map market states to portfolio weights while accounting for transaction costs, market impact, path-dependent constraints.
5. Arbitrage Machine Learning Architectures for Sparse Portfolios
-
High-dimensional portfolio selection often leads to overfitting and high turnover unless explicit sparsity constraints are enforced. This research theme designs specialized neural network architectures incorporating regularization techniques to isolate sparse, mispriced subsets of assets. These architectures are engineered to detect statistical arbitrage opportunities while maintaining investable, low-cardinality portfolios optimized mainly for execution realism.
