The role of data science in financial markets has seen considerable evolution in recent years. However, according to Stephen Ponzio, Managing Director, Head of Electronic Trading at BTIG, the fundamentals of data science and machine learning in trading have remained relatively unchanged despite the growing recognition of their value.
Speaking on the sidelines of the Equity Leaders Summit in Miami this week, he said that the use of data science, particularly in financial trading, often builds upon well-established statistical methods that have been around for a long time. Yet, the key challenge arises from the misconception that simply analyzing data can provide all the insights needed for optimal decision-making.
“First of all, a lot of the techniques that are used in machine learning, data science, or statistical analysis have been around for a long time,” Ponzio told Traders Magazine. “It’s nice that people recognize that you can learn a lot from looking at data, but the danger is thinking you can learn more than you really can.”
A prime example of this issue can be found in evaluating execution performance. Many firms strive to achieve the best prices for trades by monitoring metrics like venue analysis or spread capture. However, relying on these metrics as proxies for performance can be misleading. Venue analysis, for instance, might suggest that certain trading platforms or exchanges are superior, but according to Ponzio, “neither of those things… are correlated with performance.” In fact, the complexity of trading algorithms means that performance can’t simply be measured by looking at the source of fills.
Ponzio further illustrated this with the example of a broker’s algorithm sending orders to an ATS (Alternative Trading System). “By merely tracking the execution venue, one cannot conclude that the algorithm is performing well or poorly,” he said. In reality, the algorithm may be employing different strategies at various stages of a trade, using different venues, order types, and parameters to achieve specific goals. “Algorithms are complicated,” Ponzio explained. “They go through different phases, using different venues, different order types, different prices, and different tactics at different times.”
While data science promises to unlock new levels of insight, it is essential to understand its limitations, particularly in evaluating execution performance. Smaller trading firms, in particular, face significant challenges when trying to understand whether their executions are on track or going wrong.
Challenges for Smaller Firms
The most pressing challenge for smaller firms in trading is acquiring enough data to draw meaningful conclusions. For example, when comparing the performance of two brokers, small firms may lack sufficient order flow to distinguish between them effectively. “If you’re trying to distinguish the difference between broker A and broker B… you probably need about 5,000 orders apiece,” Ponzio noted. Smaller firms, however, often don’t have that much data at their disposal.
Beyond the volume of data, firms must also ensure their systems are configured correctly to track and interpret execution results. Without an experienced team to analyze this data, smaller firms may outsource the task to third-party companies. However, as Ponzio pointed out, simply shipping off data for evaluation might not uncover important issues. “If you’re just shipping it off to a TCA provider… are they going to go talk to the trader? Are they going to figure out why the performance is off?”
For smaller firms, one potential solution is outsourcing trading and execution to more systematic third-party providers. However, this creates another challenge—ensuring that the third-party service providers can appropriately evaluate brokers across different clients’ needs and execution strategies. Ponzio acknowledged that this is a tough problem but suggested a more systematic approach to broker evaluation in such cases.
Despite the challenges, there are promising developments in the world of crossing and algorithmic trading. One such advancement is the concept of “trajectory crossing.” This technique matches buy and sell orders at a higher level, allowing them to be executed at the average market price over a defined period, such as five minutes. By matching orders in this way, the algorithm reduces frictions and eliminates the spread cost, benefiting both the buyer and seller.
“Each side is going to get exactly the average price… it’s a gain for the algorithm on both sides ,” Ponzio explained. Trajectory crossing not only improves price execution but also helps avoid market frictions, creating a win-win for participants. “A few brokers and ATSs have already implemented trajectory crossing in their pools, further enhancing its potential,” he said.
Another promising development is the rise of private rooms within ATSs, where only selected brokers participate. This setup enables brokers to internally match their own orders, which can help mitigate market frictions and optimize the trading process. “You can have an ATS where the participants are limited to certain brokers, again, to reduce some of the frictions,” said Ponzio. Such developments are beneficial for streamlining internal trading operations and improving price execution.
The Future of Algorithmic Trading
Looking forward, algorithmic trading will continue to dominate, with the expectation that more traders will rely on algorithms to navigate the complexities of the modern market. As Ponzio succinctly puts it, “It doesn’t make sense totrade manually in the market.”
However, while algorithms are essential for efficiently navigating multiple exchanges and ATSs, there remains a role for human intuition, particularly in decisions about when and how aggressively to trade. Ponzio noted that human traders should be responsible for determining the overall strategy—when to trade, how much to trade, and the level of aggressiveness to apply. Once these decisions are made, the algorithm handles the execution.
“Problems arise when low-level decisions are controlled manually, when it’s not clear how those decisions affect performance.,” he pointed out. As algorithmic trading continues to evolve, the challenge will be in refining systems to adapt to the nuanced requirements of individual trades.
In conclusion, while data science continues to drive improvements in trading strategies, especially through innovations like trajectory crossing and private ATS rooms, the future of algorithmic trading will require a balanced approach. Algorithms will be indispensable for optimizing execution, but human oversight will remain crucial in ensuring that the overall strategy aligns with the firm’s goals.
