Abdelmadjid Laouedj has spent his career at the point where machine learning meets the messiness of real financial markets. As a quantitative researcher at a major global bank, he has built equity risk analytics and large-scale credit risk models, and his work spans factor exposure decomposition, signal extraction, and the kind of validation that separates a model that looks good from one that holds up. He trained at UC Berkeley’s Haas School of Business and CentraleSupélec in France, and he has applied his methods across insurance, oncology finance, and banking.
We spoke with him about the parts of quant finance that rarely make headlines: where standard ML approaches fall apart, how you test a model when the thing you’re predicting keeps shifting under you, and what people get wrong about how AI gets used on a trading desk. His answers cut against the popular image of AI as a price-predicting machine. The harder problem, he argues, is proving a signal is real, survives transaction costs, and works in the regimes where you actually need it.
Where do standard ML approaches break down in financial markets, and how do you work around that?
Standard ML approaches often break down when the market regime changes. A model might work well in a regular environment, a black swan event, a volatility shock, or a macro context shift can completely change market structure and correlations between assets. The way to work around this is to avoid assuming that historical relationships are, and will always be, true. I would typically test the model across different regimes or use rigorous out-of-sample validation. I can also reduce overfitting and evaluate whether the signal remains robust under different market conditions. I also think conditional approaches are important: instead of asking whether a model works “on average,” you need to be aware of when it works, when it fails, and under which conditions its predictions are reliable, so you can use the right model for the right situation.
How do you validate a model when the thing you’re predicting, market behavior, keeps changing?
You validate it by treating instability as part of the problem, not as an exception. In financial markets, the distribution constantly changes, so a simple train-test split is rarely enough. I would use out-of-sample testing, walk-forward validation, regime-based analysis, and stress testing across different market environments. If the model still performs in these different situations, then it can be considered as robust. If it has flaws, they need to be taken into account.
I would also look at whether the model’s performance is concentrated in one specific period or whether it generalizes across different conditions. The goal is not only to maximize accuracy, but to understand the stability, failures, and practical reliability of the model, so we know when to use it and when not to.
What does feature engineering look like for credit risk versus equity strategy? How different are those problems?
They are pretty different.
In credit risk, feature engineering is often focused on building metrics to measure the borrower’s ability to repay. Features may include the following: leverage, liquidity, profitability, debt structure, enterprise value, collateral, seniority, macroeconomic conditions, and industry-specific risk indicators. A good example of an industry-specific risk indicator, especially given the current focus in the market, is the balance between revenue growth and cash burn which is used for the AI/Tech sector.
The objective is usually to estimate default probability, loss severity, credit rating, or credit deterioration.
In equity strategy, feature engineering is more focused on extracting signals from market behavior and company fundamentals. These signals can include momentum, reversal, volatility, valuation, earnings revisions, factor exposures, correlation structures, liquidity, and relative performance versus peers.
So, the mindset is different: credit risk is more about downside protection and solvency, while equity strategy is more about identifying relative mispricing, alpha, exposure (to what extent a given equity is exposed to AI/Tech for example) and changing market dynamics.
How has the rise of LLMs and generative AI changed anything in quantitative finance, if at all?
LLMs have changed the workflow rather than the core principles of quant finance.
They help write code faster, help with debugging, accelerate research, and make it much easier to find relevant information by redirecting toward the right sources. Even when you do not want to rely directly on the LLM’s answer, it can still help you identify where to look, what papers to read, or how to structure a problem, more like an advanced search engine.
But they do not replace rigorous modeling. A lot of important mistakes can be found in a model developed by an LLM. In quant finance, the key questions remain the same: is the signal real, is it robust, does it survive transaction costs, and does it generalize out of sample? LLMs haven’t reached the level of sophistication necessary to make such an analysis.
You’ve worked across insurance, oncology finance, and banking. Does AI behave differently across those domains? If so, how?
The main principles are similar: you define the target, process the data, engineer relevant features, validate the model, and monitor performance/measure accuracy. But the way you validate the model and the nature of the features change significantly depending on the domain.
In financial markets, backtesting is central because you are dealing with time series, changing regimes, and live market behavior. In credit risk, validation is more focused on default events, calibration, and regulatory expectations. In oncology finance, the problem is less about market prediction and more about operational forecasting, cash-flow planning, and decision support, which the structure can stay more or less constant over time.
In insurance, models are often centered around pricing, claims frequency and severity, and long-term risk estimation. There is a strong emphasis on stability and regulatory compliance. So AI does not “behave” differently in a mathematical sense, but the data, constraints and validation framework are very different across these domains.
What’s the biggest misconception people outside quant finance have about how AI is actually used your space?
I’d say the biggest misconception is that AI is used as a black box that simply predicts the market.
In reality, most serious quant work is not about asking a model to magically forecast prices, without any serious analysis. It is about building carefully tested signals, understanding their economic rationale, and validating whether they remain useful after costs, constraints, and regime changes. Most of the work lies in the preparation of the data, the choices of the metrics, the backtesting, the interpretation of the output, etc
A model that looks impressive in sample but fails out of sample is not useful. In quant finance, the hard part is not just building a complex model; it is proving that the signal is robust, implementable, and not just some noise.
