Process & Data Collection
Primary research through interviews with quantitative practitioners and academics, paired with a synthesized literature review.
Methodology
In order to supplement the findings from the literature review, primary research was conducted through interviews. Interviews were selected as the primary method because they allow for deeper insight into how ML is actually used in the field, especially in trading spaces where execution strategies and decision making while trading require context and aren\u2019t easily documented through quantitative surveys.
Interviews
"Execution quality is constrained by liquidity, volatility … and how the market reacts to orders, not just by if the model predicts direction"
"Ai is more there as a support role… it’s helping you build processes and tools"
"You can have the best model in the world… but if you don’t know how to trade it.. You aren't going to make money"
"Financial markets are highly noisy… models that perform well historically may fail… in changing conditions"
"The human still needs to have a framework in which to evaluate the outputs of AI systems"
Cross-Interview Analysis
Across all the interviews, there was an extremely strong consensus that ML and AI systems are valuable tools within institutional trading, but also that they are not capable of independently improving execution quality, and that they are indeed limited by market microstructure constraints. Nearly every respondent emphasized that the execution quality of a trade heavily depends on external factors like liquidity, volatility, and order flow instead of just on predictive accuracy. A major pattern that constantly appeared throughout the interviews was the comparison between signal generation and execution. Both Mr. Veeramani and Mr. Cheng explained that ML models are mostly useful for identifying statistical trends and generating signals based off of those trends, as well as improving operational efficiency. However, they also constantly stressed that the actual execution of trades, and therefore the quality of these trades, is constrained by ever-changing conditions and human decisions.
Another major trend in the interviews was the limitation of AI systems in dynamic environments. Dr. Dixon explained that financial markets are always evolving which causes many ML models to overfit historical data and fail when conditions shift. Similarly Professor Kalodimos also discussed the importance of creating conceptual frameworks to evaluate trades before applying ML systems, showing how AI outputs are only useful when humans are capable of interpreting them. Overall, these findings strongly supported my original hypothesis that while ML systems are able to aid in institutional trading environments through predictive modeling, execution quality remains constrained by market microstructure conditions.