An interview with Olivier Gandouet, Vice President & Director, Quantitative Equity at TD Asset Management

There are many challenges that arise when integrating machine learning-based outputs within investment decisions, including:

1. Data Acquisition and Cleaning: Managing very large amounts of heterogeneous data sources, each with its own share of errors (revisions, poorly reported data, non-homogeneous, unreliable sources).
2. Heteroskedasticity of the Market: Market moments are often unique and can yield outcomes that could not have been learned from past patterns.
3. Natural Decay of Signal: The competitive nature of the finance field means that the quality of stock assessments varies with the average improvement of the market.
4. Over-Parametrized Models: Highly over-parametrized models give machine learning solutions very high in-sample fitting potential.

All these points combined necessitate a crisp interpretability of the signal, which is even more critical than for any other form of model. Since signal capture can emanate from a vast dimensional field, interactions and intricacies can be hard to fully understand, even when leveraging model explanation frameworks (SHAP, LIME, or Integrated Gradients). More recently, people have tried to mitigate the limitations of model explanation tools by using large language models (LLMs) to textually rationalize the signal. However, there is still little evidence that convincing textual conclusions will materialize in abnormal returns (it is easy to find good explanations for phenomena that will not realize).

The amount of reasonable information that can be used to make informed decisions in the market is known to be quite low (“efficient market”). Most captured patterns by these powerful fitting machines can tie to underlying risks and yield very poor risk-adjusted returns when the statistical bond that explained the captured pattern breaks. To partially address this fundamental problem, quantitative asset managers need to adapt machine learning solutions to the reality of the market and add high levels of regularization (from classical L2-L1, dropout, causal frameworks, or risk neutralization) that go beyond out-of-the-box machine learning solutions. This requires a high level of expertise in mathematics, machine learning engineering, and quantitative finance, which is still a niche intersection. Making good adaptations is both hard to achieve and has potential.

Another major hurdle of complex machine learning pipelines can come from deployment. The most advanced versions can lead to high compute times or complex data ingestion pipelines that require extreme care in each step to ensure that all data distributions are as expected. Delivering a model with a high level of validation can sometimes require almost the same amount of time (if not more) as creating it in the first place.

Lastly, a phenomenon that has arisen quite recently (post-ChatGPT release) is the need to align solutions at the same pace that innovation is driven in the field. With new frameworks and processes deprecating very recent advancements at an unprecedented pace, architecture choices have become more critical than ever and force longer design cycles, even if the development cycle has been shortened significantly with tools like Copilot or Cursor. We feel that the biggest bottleneck today comes from the validation of output by informed experts.How do you ensure the transparency, explainability, and regulatory readiness of ML-driven signals used in portfolio construction?

What strategies have proven effective in embracing collaboration between quant researchers, data engineers, and compliance or risk teams within TD?

Quantitative equity strategies were among the first to integrate AI tools, building on existing knowledge and understanding of machine learning and technical tools. However, even for fundamental researchers, the rise of generative AI (GENAI) tools like search, which are relatively simple to integrate, has significantly disrupted the ability of fundamental analysts to cover a larger span of companies across wider datasets. (Beware of market concentration: even if you go faster, if everyone drives in the same lane, the risk of crashes drastically increases.)

What types of infrastructure or tooling investments are you prioritizing to support scalable and stable ML adoption in portfolio research?

Cloud infrastructure is paramount for fast access to tooling and scalability. Nevertheless, we believe that on-premises or local machines can still be the way to go to mitigate the increasing cost of computing, ensure robustness to outages, and control the cost of some of the compute required by this type of machine. Why pay for token scoring if 80% of your scoring can be done nearly identically in a more controlled and cost-effective way?

Data acquisition and creation have become even more important than before. The ability to purposefully navigate and integrate deep information from news and media has necessitated acquiring large amounts of point-in-time data. This allows for rationalizing signals that could be inferred from recent events, providing a form of backtesting on point-in-time information.

Recruiting individuals knowledgeable in AI has become an increasing priority in the investment field. This ensures that we have the expertise to build, validate, and deploy at a pace that keeps us ahead of our competitors.

In terms of our event format, (our conferences are informal and intimate peer-led meetings where all speakers and delegates are senior executives from top financial institutions), how do you see it assisting you with overcoming the challenges faced?