Interpretability methods comparison
What do different explanation methods reveal about the same model?
SHAP, LIME, PDP and ICE applied to a California housing model to examine global patterns, individual predictions and contextual variation.
ReadyProjects
Current work on credit decisions, synthetic financial data and portfolio allocation.
Completed experiments and earlier applied projects sit below.
Credit decisions under constraints
Can a credit decision be statistically defensible, economically rational and legally accountable?
A decision system where prediction is only one component. Regulation belongs inside the modelling problem, not as a layer added afterwards.
Synthetic financial data
What does it mean for synthetic financial data to be statistically and financially faithful?
Generation is only half of the problem. Synthetic series must preserve the statistical structures that make financial data meaningful for inference and risk.
Portfolio allocation under uncertainty
How should capital be allocated when expected returns, risks and dependencies are themselves uncertain?
The mathematically optimal portfolio for estimated parameters may not be a robust investment decision. Parameter uncertainty therefore becomes part of the allocation problem itself.
Comparing PCA, PCR and PLS
When does supervised dimensionality reduction outperform variance-based representation?
An empirical comparison separating the representation and prediction objectives of PCA, PCR and PLS.
CompletingReproducing unemployment duration models
Can a published duration analysis be recovered from its stated specification and data construction?
An econometric reproduction of unemployment durations in Austria, focused on specification, data construction and reproducibility.
CompletingNon-parametric estimation
How much structure can be recovered without specifying a functional form?
Kernel density estimation and Nadaraya–Watson regression, with bandwidth selection as the central problem governing smoothness, bias and variance.
DraftingDigital marketing adviser
Ipsos- Decision
- Agentic Systems
3D obstacle detection
Airbus Helicopters- Perception
- Geometric Learning
Animal behaviour analysis
CENTURI Living Systems- Perception
- Kinematics
California housing
Aix-Marseille School of Economics- Prediction
- Explainability
Market regimes and dependence
Aix-Marseille School of Economics- Modeling
- Dependence
Portfolio allocation under uncertainty
How should capital be allocated when expected returns, risks and dependencies are themselves uncertain?
The mathematically optimal portfolio for estimated parameters may not be a robust investment decision. Parameter uncertainty therefore becomes part of the allocation problem itself.
How much estimated optimality should be sacrificed for robustness?
The experimental starting point is to demonstrate the problem: estimate parameters, optimise, then perturb the estimates and observe changes in weights, concentration, turnover, expected risk and out-of-sample behaviour. A central diagnostic is that a small change in assumptions can produce a large change in allocation.
Baselines are included as different responses to the allocation problem — equal weight, mean-variance, minimum variance, risk parity, maximum diversification, CVaR — not as a collection of algorithms. Evaluation is planned to include volatility, drawdown, CVaR, turnover, concentration, costs and allocation stability, not only return and Sharpe ratio.
Regime-aware allocation is a possible later extension, once the core allocator is rigorous. The project has not yet produced allocation results.
Planned Dimensions / Sous réserve d'investigation
- Estimation error
- Allocation baselines
- Robustness
- Evaluation beyond Sharpe