The year 2025 represents a pivotal milestone in the evolution of artificial intelligence (AI), characterized by the emergence of advanced models such as DeepSeek, which signify a new frontier in AI capabilities. These developments further substantiate the viability of open-source AI models, underscoring their potential to drive innovation and enhance efficiency in this transformative era. This trajectory is expected to have a profound and enduring impact on human history, particularly in industries that are heavily dependent on data.
For financial professionals, a key question arises: has AI already impacted the industry, and if so, what fundamental changes will it bring? One of the most immediate effects is seen in the buy-side investment sector, particularly among private equity (PE) firms and hedge funds reliant on fundamental analysis. AI models like ChatGPT and DeepSeek now outperform human analysts in processing and interpreting financial reports, sparking significant debate on their implications for investment strategies. However, an equally significant yet less visible transformation is occurring in trading. This study examines the applications of AI in the trading sector and evaluates their impact on the business models of both sell-side and buy-side participants.
Fundamentals and definition of AI in the trading context
The definition of artificial intelligence (AI) remains ambiguous and is often conflated with recent advancements in deep neural networks, particularly transformer-based models. While these models represent the current state-of-the-art, they should not be regarded as the definitive path to achieving general artificial intelligence (AGI). A relevant example is the author’s experience in developing a chatbot in 2015–2016, when Long Short-Term Memory (LSTM) networks were the dominant approach; however, LSTM models have since fallen out of favour. This progression highlights the rapid evolution of AI technologies and suggests that the pursuit of AGI will likely involve a continuous cycle of innovation and methodological advancements.
In the context of trading, AI can be defined as a process in which a model is trained on historical financial data to analyze market patterns, optimize strategies, and support decision-making.
An intuitive representation of AI in trading is illustrated in Figure 1. In this framework, trading activities generate data, which is recorded in post-trade transaction cost analysis (TCA) and subsequently used to train a pre-trade model that informs future trading decisions. This process closely resembles reinforcement learning and is sometimes referred to as data augmentation or electronic trading, depending on the sophistication of pre-trade model and level of automation. To ensure clarity and consistency, we will refer to this concept as AI for trading throughout this article.
Once we establish a clear definition of AI in our trading context, it is crucial to acknowledge that the advancement of AI generally depends on three main factors, as illustrated in Figure 2.
Computation: Although the size of transistors is physically limited by the quantum tunneling effect, we can continue to adhere to Moore’s Law by transitioning to 3D structures through advanced packaging systems, such as CoWoS (Chip-on-Wafer-on-Substrate) at TSMC. This approach suggests that the increase in computational power can persist into the foreseeable future, mitigating concerns about computational resources reaching their limits.
Model: There is significant investment in model development by various entities, including OpenAI, DeepSeek, Mistral, Anthropic, and Meta, among others. The field is characterized by rapid iteration, with innovations emerging on a daily basis.
Data: Data availability poses the most significant concern regarding the scalability of AI. There is a looming risk of exhausting the open data available on the internet, which could become the most substantial bottleneck to AI’s progress. While some argue that reinforcement learning might address some data challenges, it does not fundamentally generate new information.
In terms of advancement, governments and high-tech firms have invested billions of dollars in model development and GPU clusters. If computational resources are viewed as infrastructure—akin to highways accessible to society at large—then the progress of AI in the financial industry is inherently linked to advancements in models and the availability of trading data. Our forecast of AI transformation will be based on recent developments in data and models, as these factors are key drivers of AI-driven innovation in financial markets.
Short term change
Before the introduction of MiFID, there was limited shared trading data between the sell-side and buy-side for foreign exchange (FX) transactions. This data was originally stored in disparate formats, with varying metric definitions across different banks and platforms. Such data fragmentation posed a significant obstacle to the advancement of the FX industry.
For example, prior to MiFID, evaluating and understanding black-box algorithms across banks was nearly impossible. Reflecting on my experience as a quantitative analyst at Morgan Stanley a decade ago, this period was particularly challenging. Despite efforts to introduce new algorithmic trading strategies, TWAP (Time- Weighted Average Price) and VWAP (Volume-Weighted Average Price) remained the most widely adopted among clients. In the pre-MiFID era, decision-making was predominantly controlled by senior traders on the desk, who prioritized simplicity and transparency. This preference contributed to the sustained dominance of TWAP and VWAP strategies.
However, with the introduction of MiFID and the emergence of transaction cost analysis (TCA) a significant shift has occurred in the FX trading industry. For the first time, as illustrated in Figure 3, a comprehensive common database, encompassing $800 trillion USD in transactions, has been established with standardized definitions of various performance metrics. This unified data infrastructure has laid a solid foundation for advancing AI in the trading community, enabling more sophisticated data-driven strategies and automation in financial markets.
If we categorize trading AI systems that rely on basic statistical methods and still require human interpretation as Level 1 AI (L1 AI), the evolution of AI in trading can be analogized to advancements in autonomous driving: progressing from a human driven level (L0) to data-driven (L1 AI), and ultimately toward driverless Level 4 AI (L4 AI). This transition signifies a gradual reduction in human intervention, mirroring developments in the automotive industry, where the objective is to achieve fully autonomous systems.
In this section, I aim to demonstrate that data-driven Level 1 AI (L1 AI) is already widely adopted in the FX trading industry. To further analyse this transformation, I propose the following two hypotheses:
H0: FX trading is still predominantly Human Driven Level 0 (L0), which would support the continued popularity of traditional algorithms such as Time-Weighted Average Price (TWAP) and Volume-Weighted Average Price (VWAP).
H1: Data Driven L1 AI has already been adopted on a large scale within the FX industry, which would lead to the increased popularity of algorithms that outperform traditional approaches in execution efficiency and performance
From Figure 4, it is evident that sophisticated opportunistic algorithms, such as arrival price algorithms, account for approximately 60% of algorithm usage, whereas TWAP and VWAP combined represent only 5% of algorithm usage.
This is supported by Figure 5, which demonstrates that opportunistic algorithms exhibit the best risk transfer price performance compared to other algorithmic trading styles.
Therefore, it is reasonable to reject the hypothesis H0, which posits that human decision-making dominates the execution industry, and instead accept our assumption H1, confirming that data-driven Level 1 AI (L1 AI) has already been implemented in the FX execution space.
Another significant trend to highlight is the ’winner-take-all’ effect. It is anticipated that algorithm providers who deliver superior performance will attract more business, putting pressure on the sell-side to enhance their algorithm performance. This dynamic is expected to further accelerate the shift towards more sophisticated and effective trading algorithms within the industry.
All the transformations discussed, though straightforward and intuitive, would not have been feasible without the existence of data. Since the implementation of MiFID, TCA providers have played an indispensable role by aggregating trading data and offering analytical tools. These providers go beyond merely facilitating regulatory compliance; they are crucial in enabling the data-driven transformations observed across the industry.
Given the three foundational elements of AI—Data, Computation, and Model—the anticipated expansion of TCA’s role is profound. TCA is expected to become increasingly vital in the AI-driven evolution of the trading industry, serving as a cornerstone that supplies both data and models to AI systems.
Medium term change
As the volume of data naturally increases, as shown in Figure 3, it necessitates enhancements in model sophistication to achieve better execution outcomes. This progression is pivotal for advancing from Level 1 AI (L1 AI) to Level 2 AI (L2 AI) in trading. This evolution underscores the need for models to adapt and evolve in complexity, aligning with the increasing granularity and volume of available data, thereby enabling more effective and efficient trading decisions.
As shown in Figure 6, the usage of algorithms as a percentage of RFQ spot volume has remained around 50%, with a slight upward trend recently.
The fluctuations in algorithm adoption can largely be attributed to the lack of decision-support models that help buy-side participants determine whether to use RFQ or algorithms based on real-time market conditions—an approach known as the RFQ vs. Algo model.
While sophisticated high-frequency trading (HFT) firms may have developed such models, the majority of FX market participants likely lack the resources to build and implement these advanced decision-making systems. This resource gap limits broader algorithm adoption, as RFQ remains a more accessible and lower-risk option for many traders.
As illustrated in Figure 7, algorithms tend to outperform RFQ in terms of slippage, particularly for larger trade sizes and in less liquid markets.
This performance suggests that algorithmic trading has effectively supplanted the traditional role of
voice trading, which has historically been considered superior for handling illiquid and large transactions. Consequently, the primary competition for traditional voice traders no longer comes from electronic principal trading desks but rather from algorithmic trading desks.
If such models become accessible through multi-dealer platforms or Transaction Cost Analysis (TCA) providers, the adoption of algorithmic trading is expected to increase. According to the BestX paper Li (2025), FX volume is inversely correlated with the Federal Reserve’s decision-making cycle, and given the current phase of decreasing interest rates, it is reasonable to anticipate a rise in algorithmic trading. This increase is expected both in terms of its percentage relative to RFQ volume and in absolute trading volume.
In abstract terms, the author posits that the transition from **Level 1 AI (L1 AI) to Level 2 AI (L2 AI)** in trading is driven by advancements in trading decision models, such as the RFQ vs. Algo model. Each progression in this domain is expected to fundamentally reshape the current trading landscape.
Vision for long term future
While forecasting future trends presents an inherent challenge, it remains possible to outline the probable trajectory of FX trading or financial trading in general.
Customization is the key direction. Best execution has different implications for different clients. Consider a scenario in which a trader needs to quickly execute several large transactions due to international MA deals; hence, he wants to assign 70% weight to signaling risk and 30% to the risk transfer price. Meanwhile, another short-term trader who aims to capitalize on short-term momentum in EUR/USD assigns 100% weight to the risk transfer price. This example illustrates that the best execution or utility function is fundamentally different for different clients based on the nature of their flow.
As illustrated in Figure 8, buy-side traders assign different weights to various benchmarks in their Transaction Cost Analysis (TCA) performance evaluation and algorithm selections. In the traditional setting, designing a customized algorithm to achieve the Best Execution for individual strategies based on the nature of their flow is nearly impossible. However, with advancements in artificial intelligence (AI), particularly recent developments, the creation of bespoke financial products—such as algorithms designed to optimize a bespoke benchmark— will become crucial for future financial product offering.
AI as a Copilot in Trading.Copilot mode, i.e., transitioning from Level 2 (L2) to Level 3 (L3) automation, will take a considerable amount of time to become a standard in the financial industry due to regulatory complexity and the risks associated with tail events. Similar to developments in the autonomous driving industry, achieving Level 4 (L4) AI in financial trading is unlikely in the foreseeable future. Instead, within the algorithmic trading community, AI will primarily function as a copilot—providing analytics and assisting traders rather than replacing them. Consequently, the ability to leverage data and analytics from multiple sources will be a crucial skill for all market participants.
The Tipping Point of AI Adoption.There exists a tipping point at which change occurs rapidly. Historically, significant technological advancements remain subtle for long periods before suddenly transforming industries. For example, deep learning was not widely recognized until the success of AlexNet in 2012, after which machine learning models converged on deep learning techniques.
Similarly, the emergence of ChatGPT has captured global attention. Although the financial industry may not yet feel the full impact of AI, preparing for the AI Age is imperative. When widespread adoption occurs, it will happen swiftly, with a winner-takes-all effect.
Conclusion
The change is already underway, even if it remains subtle at present. This transformation will be fundamental for all professionals working in finance. A necessary investment in data, models, and computational resources is imperative for both buyside and sell-side participants. As Sam Altman recently stated, those who fail to adapt risk being on the wrong side of history. In conclusion, I have outlined short, medium, and long-term changes for FX trading. With advancements in TCA, data, and algorithmic offerings, achieving best execution will become increasingly feasible.
Finally I want to present an open question for all readers: If we achieve Level 4 (L4) automation, where all financial decisions are made by AI, does our civilization still truly belong to us?
