Abstract

Trading has been revolutionised by the rapid adoption of AI-driven trading models. These AI-driven trading models hope to enhance decision-making through improving speed, data-processing and predictive abilities. However, there have been debates regarding the usage of these trading models and their possibility in magnifying systemic risks during periods of market pressure. This paper explores to what extent do AI-driven trading models improve trading decisions without increasing systemic risk during periods of market volatility. Through in-depth analysis, we evaluate how effectively AI-driven trading models perform during times of market volatility and explore the mechanism through which AI makes improvements in decision making in trading (e.g., speed and forecasting) and through which it can amplify systemic risk (e.g., herding). The discussion considers whether the benefits of AI-driven trading models outweigh its potential costs. Our findings provide a deep understanding of the benefits and risks of AI-driven trading models used in decision making and how to manage and mitigate potential risks.

Introduction

Trading in financial markets is extremely risky and complex due to external economic and psychological factors. This is why economists deployed the rapid use of AI-driven trading models to help the investor make better trading decisions or carry out trading automatically. 60-80% of equity trading volume is accounted for by AI-driven algorithmic trading, and approximately 91% of hedge funds report using or planning to use artificial intelligence (AI) tools. Although the integration of AI promises a revolutionary and unprecedented improvement in trading efficiency, such as accelerating processing of large volumes of market data, strengthening forecasting capabilities, leading to adaptive trading strategies and more, it is a double-edged sword. These same systems can lead to instability, particularly during periods of market volatility when their behaviour is most consequential. 

The aim of this research paper is to investigate that core tension, and the sustainable adoption of AI in trading. This paper examines the intersection between AI’s potential and systemic risk, specifically in the context of market volatility. As AI is quickly becoming ubiquitous in global markets and systemic risks are a borderless threat (a shock in one market can quickly spill across borders), studies on how to employ the use of AI while balancing its risks are necessary due to the high stakes it has on global financial stability.

Scope and Structure

Part I: The Benefits of AI in Trading Decisions examines the systems through which AI-driven trading models help improve trading decisions by improving efficiency and speed.

Part II: The Danger of Amplified Systemic Risks examines the mechanisms by which AI-driven trading models can destabilise markets during periods of volatility. 

Part III: Regulating the AI-driven Financial Market combines the findings and evaluates whether AI’s potential and recorded benefits outweigh the overall risk of using such a programme.

Literature Review

Numerous reviews have examined the benefits and risks of AI in trading.

For instance, Gong et al’s (2026) synthesis discovered that in asset pricing and credit access, advanced machine learning methods enhance predictive accuracy, although they introduce challenges in algorithmic fairness, interpretability and new vectors for systemic risk such as algorithmic collusion. 

Hansen and Lee (2025), using laboratory-style experiments, discovered that LLM-based AI models made “more rational decisions than humans, relying predominantly on private information over market trends”. 

Qiu and Han (2026) created a structural multi-agent model which formalises “representation homogeneity” and theoretically proves that AI agents can converge under stress due to shared information coding even when they seem to make different decisions during calm conditions. 

Zhang (2025) studied an extension to the Capital Asset Pricing Model that would untangle systemic risk, implying that an extensive adoption of common foundation models creates a new source of systemic risk.

Sabkha and Jbir (2025) gave empirical evidence that “AI adoption does not increase herding on average; however, higher aggregate AI exposure significantly amplifies nonlinear herding under adverse market conditions”.

Methodology

This study employs the use of qualitative analysis of the mechanisms through which AI-driven models improve trading decisions and amplify systemic risks, through textual analysis of existing academic research papers on this topic. 

Analysis

Part I: The Benefits of AI in Trading Decisions

The first aspect to be examined is how AI’s capacities are able to assist in day-to-day activities and speed them up substantially, saving time and resources for trading companies. Firstly, AI is able to process large volumes of market data more efficiently and thoroughly than any human, simultaneously saving time and finding more revelations from a particular analysis that most traders would have most likely missed. For example, as Interactive Brokers (IBKR) features AI tools such as Toggle AI, which sifts through billions of data points every day (Interactive Brokers, n.d.), this heavy use of AI allows IBKR-based investors to speed up operations and research, allowing them to make smarter and more in-depth choices when managing their portfolios. Additionally, artificial intelligence can improve forecasting and decisions through machine learning (ML). ML’s usefulness appears during the processing of massive datasets with millions and millions of data points, where such an activity becomes nearly instantaneous (Aldasoro et al., 2024). This approach allows for the uncovering of hidden patterns, as well as being able to automate complex choice processes. Machine learning allows AI to not only be faster but also more accurate than humans, thus improving trading decisions as a whole.

The next avenue of analysis to explore is how ML and reinforcement learning (RL) both work hand-in-hand to improve and assist traders. As discussed above, ML’s usefulness derives from the capacity of consuming large amounts of data to perceive different situations, market positions or repeating patterns and thus draw increasingly accurate predictions (Aldasoro et al., 2024). Reinforcement learning is a core branch of ML; unlike ML however, RL’s usefulness is due to a repetition of trial-and-error type situations made by artificial intelligence (Li, 2021). Although RL is by definition a subtype of ML, it remains crucial in providing primary data through testing itself, as well as enabling trading strategies to adapt in real time. Furthermore, AI’s usefulness can also be found on the financial side of trading: more specifically, the decrease in expenses required for human resources (Aldasoro et al., 2024). By allowing AI to replace certain roles or greatly speed up processes, large companies can reduce costs on salaries and training and overall improve their time-to-money ratio. The combination of self-improving AI and lower costs shows how trading and decisions are greatly improved by AI.

Moreover, these same efficiency and decision-making gains are tested most severely not in stable markets, but during periods of volatility. In situations of trading market volatility, these periods with potentially increased risks and benefits are characterised by prices changing with extreme rapidity and wildness. In essence, periods of high market volatility can be make-or-break for trading companies and require immense skill and quick and accurate thinking to succeed. On a day-to-day basis, mild to high market volatility is considered very normal and occurs relatively frequently, however extreme panic-driven spikes are much more rare. To understand AI’s role in improving enterprises during these periods, we must first explore why market volatility occurs and how financial success can be achieved in such times. Through research, we find that the main causes often come from shifting interest rates, bad economic news, potential wars or conflicts, as well as specific political events (HSBC, n.d.). Such an issue can be approached one of two ways: either through a historical approach by using past data, trying to find similarities or patterns within the data to base decisions upon, or through future predictions based on knowledge, feelings and utilising tools such as the CBOE Volatility Index (VIX). AI’s main benefit in such market conditions is its overall speed. This speed can be derived into two main components or subsections: research and trading activity. On one hand, AI is able to seek out information, identify patterns and link data between two or more variables much faster than any human. On the other hand, AI’s speed also lies in how quickly it can respond to market movement. AI and automated trading systems are vastly faster than human traders, executing trades and orders in microseconds compared to the average human’s 0.4 seconds (Tasdelen, 2026).

Part II: The Danger of Amplified Risks

The very same speed – the one that seems so appealing and beneficial, and the rationale behind the widespread adoption of AI in trading – becomes a liability when unique situations appear. Unique or market behaviours occur due to unforeseen or unaccounted for metrics that impact the market. Certain examples include the terrorist bombings in 2001, the 2008 financial crisis or, even more recently, the COVID-19 pandemic. Although completely different events in their own regard, these elements all have one thing in common; they are all unique and unforeseen. In more precise trading terminology, such events are called “black swans” (Investopedia, n.d.) and usually occur due to recurring causes such as overall market complexity or even interconnected global systems driven by human psychology. Such events cause a great amount of shock, confusion and terror due to their improbable occurence outside of normal statistical models.

When considering AI, black swan occurrences can be dangerous (Danielsson and Uthemann, 2023, cited in Aldasoro et al., 2024). In effect, data limitations can be detrimental for companies using AI during unprecedented events. As market volatility increases, certain programmes or AI systems might break down or crash as a result of improbable circumstances. As AI-driven models rely on historical data for training, when a black swan event occurs, these models “don’t just underperform—they break”, as noted by Grigory Chikishev, a quantitative trader, who also observes that the expectation of a machine learning model to predict such a peculiarity is “mathematically flawed”. As a whole, black swan events can not only break or cause issues regarding the AI itself, but can also lead to significant financial losses for its parent companies. This creates a paradox: although AI can respond to changing market conditions with amazing speed, it cannot understand an unfamiliar shock.

Beyond the systemic risk and opacity concerns already discussed, additional factors limit how far AI should be implemented in investment bank trading.

Firstly, the high cost of error. Using AI for trading and lending decisions carries a huge risk of expensive mistakes, especially since these systems mean humans’ involvement carries less weight. If a poorly calibrated or malfunctioning AI executes trades or approves exposures at scale, the resulting financial damage could be substantial (Maple et al., 2023). Furthermore, this risk worsens because of the opacity problem; an error hidden in a model’s internal logic might not show up until it has already caused losses.

Secondly, is the lack of accountability. The drawback lies in who to blame if the AI makes a mistake. Machine learning models update constantly based on new data. When a critical decision turns out flawed, it is often unclear whether the fault lies with the training data, the model architecture, how it was deployed or the humans who approved it (Maple et al., 2023). This ambiguity creates governance dilemmas. Financial institutions are used to clear accountability for trading decisions. The use of AI makes it difficult to hold someone accountable and improve from the mistake.

Thirdly, data privacy and data quality challenges also limit the effectiveness of AI in investment bank trading. AI trading systems need access to large amounts of data. Financial institutions have historically guarded this data tightly because of its commercial value and strict regulations like the GDPR. Poor data quality (i.e., incomplete, biased, inconsistent) can directly wreck model performance and produce unreliable trading signals. Additionally, pulling data from multiple sources to train these models introduces its own security and compliance risks (Maple et al., 2023). A model trained on unrepresentative historical data might work well under normal conditions but fail exactly when conditions deviate from what it has learnt, which could be a decisive factor in volatile markets.

Moreover, the interaction of AI models at the market level can lead to systemic risks that jeopardise the financial market. There are two primary ways through which this can occur: algorithmic herding and increasing market concentration.

Algorithmic herding is when AI models, instead of acting separately, coincide on similar investment decisions. AI systems do not “irrationally” follow the crowd, which is the case in information herding, but independently converge to the same conclusion as they analyse similar information sources (e.g., macroeconomic data and financial news) through alike or shared foundation models. Algorithmic trading amplifies volatility as it converts isolated analytical errors or small market triggers into large-scale and system-wide problems. Coinciding AI responses have been linked to real-world events, such as the August 2024 Nikkei collapse (a 12.4% single-day decline after a modest rate change) and the immediate 67% Lyft after-hours surge triggered by a typographical error. 

Additionally, an increase in market concentration creates a “monoculture” risk, meaning that AI systems become more fragile as diversity of responses is reduced. For example, 90% of advanced processors are produced by one company, Taiwan Semiconductor Manufacturing Company (TSMC), and the global cloud capacity, the total computing power, storage volume and network throughput, is controlled by just a few firms. This creates a new vector for systemic risk as a single point of failure (e.g., cyberattack on a cloud server) or a bug in a popular foundation model can affect the entire financial system, as many firms rely on the same fundamental infrastructure. 

Job displacement and deskilling is another disadvantage for AI in trading. Bloomberg Intelligence projects Wall Street banks could cut up to 200,000 jobs over the coming years due to intelligent automation, with junior associates and entry-level analysts most exposed given the routine nature of much of their work (Paulise, 2026). Beyond direct job losses, researchers worry about deskilling. As AI takes over more analytical and decision-making work in trading, people working alongside these systems may have fewer chances to develop the independent judgement that used to come from hands-on experience (Maple et al., 2023). This creates a longer-term structural risk which is easy to overlook when firms are focused on immediate efficiency gains. If fewer people develop deep trading expertise because AI handles routine analysis, the pool of experienced workers who can spot when an AI system is behaving unusually – and intervene properly – may shrink over time.

Finally, regulatory lag is a cross-cutting limitation: regulatory frameworks have struggled to keep up with the speed at which AI is being adopted into trading. Bodies like the EU (through its AI Act) and India’s SEBI have introduced specific requirements around registration, logging and human oversight of algorithmic trading systems, but these frameworks are recent, and their practical effectiveness remains largely untested against a genuine AI-driven market crisis (Hilton, 2026). This lag is important because many of the disadvantages discussed above (e.g., cost of error, accountability gaps, data risk) are exactly the issues that well-designed regulation is meant to address.

As a whole, these limitations reinforce a running theme. The disadvantages of AI in investment bank trading are not isolated technical quirks, but factors that interact: opacity makes errors harder to detect and correct, deskilling reduces the human capacity needed to catch those errors when models fail, and regulatory frameworks are still catching up to a technology whose risks compound across several dimensions at once, rather than presenting as a single, containable problem.

Discussion

Part III: Regulating the AI-Driven Financial Market

The literature reviewed above considers two well-known risks: amplification through the idea of herd mentality and reduced transparency as a result of black-box systems. It also contains an unresolved question: how seriously should these risks be taken? The discussion below talks about the convergence mechanism, tests it against historical events, mentions how “model opacity” increases risk and weighs the strongest counterargument against it.

The convergence mechanism, examined

The systemic risk argument is based on a specific chain. AI models trained on very similar data produce similar outputs, ultimately changing the decisions made in trading. These decisions reduce the change in market behaviour. Less change means a shock hitting one strategy hits many at once. This creates more of a risk, rather than preventing it. Danielsson et al. (2021) and Svetlova (2022) lay out the theory. Meng and Chen (2026) show that similarity among institutional managers has gone up with AI adoption. Notably, this is not just an opinion-based pattern: similarity increased the most during the COVID-19 crash. This is exactly when the herding mechanism should be at its worst and have the largest, most negative impact. This means the theoretical argument now carries more weight than merely a speculative thought based on the evidence from Meng and Chen.

Despite this, caution is needed about overuse of evidence. Similarity with large managers will increase for reasons which are unrelated to AI adoption, including the use of “index-tracking”, pressure from regulation or simple market consolidation. The literary review does not rule these out, therefore it is a drawback of the evidence rather than a flaw in a study. The theoretical mechanism (correlated signals from shared training data) is well specified and plausible, and the empirical trend is consistent.

Historical precedent versus present-day capability

The 2010 Flash Crash is evidence for AI-driven systemic risk, though it warrants some care. The algorithms in 2010 were considerably simpler and worse compared to the deep learning and reinforcement learning systems now being discussed. The flash crash in 2010 shows that automated trading can cause severe market crashes; what it does not show is that more advanced AI will produce dislocations through the same mechanism or at the same scale. The better question is not whether automated trading can cause crashes but whether the specific properties of more modern AI make these events more frequent or more severe than earlier trading. The literature suggests this is likely, however  it remains a reasonable projection from established mechanisms rather than something directly observed at scale in a genuine crisis.

The opacity problem as a compounding factor

The black-box problem does not operate independently of systemic risk; it compounds it. If AI-driven strategies were highly correlated but fully transparent, risk managers could identify the correlation and take steps to reduce it, such as diversification requirements, position limits or manual override. The core difficulty Cohen, Snow and Szpruch (2021) identify is that this kind of intervention becomes much harder when the models generating correlated behaviour cannot be readily interrogated. A risk manager might observe portfolios converging without knowing why or which specific model features drive the convergence, which limits the ability to design a targeted fix. The point cuts deeper than simply noting that AI models are hard to understand. It identifies a specific way opacity undermines an existing risk-management practice, profit-and-loss attribution, that depends on explaining, not just observing, trading outcomes.

The historical parallel to pre-2008 copula models are instructive here, though it should not be overstated. The 2007-2008 crisis is often cited as an example of financial models becoming dangerously disconnected from the risks they were meant to represent. Cohen, Snow and Szpruch argue today’s ML models pose a similar or greater risk. The comparison is useful for illustrating the general danger of over-relying on models whose internal logic is poorly understood. Yet the specific failure mode in 2008, systematically underestimating correlated default risk in a way the models were structurally unable to capture, is not identical to the failure mode most discussed in the AI trading literature: correlated trading behaviour across many independently deployed models. The two share a family resemblance in that complexity outpaces comprehension in both cases, but the precise mechanisms differ and conflating them risks overstating how directly one historical episode predicts the other.

Weighing the counterargument

The strongest counterargument in the literature is survey evidence that most financial institutions currently favour simpler supervised learning models over complex deep learning or reinforcement learning systems, and that most firms maintain significant human oversight over trading operations (Sidley Austin LLP, 2024, citing De Nederlandsche Bank and AFM survey data). This is a meaningful qualification, since much of the systemic risk and opacity literature is, to some extent, discussing a risk that is more prospective than current, based on where AI adoption in trading appears to be heading rather than where it currently sits for the median institution.

This counterargument nonetheless has two limitations worth noting. First, survey-based evidence reflects self-reported practice at the time of the survey and can quickly become outdated in a fast-moving technology area; the same institutions surveyed may already be piloting more advanced systems not yet reflected in aggregate adoption figures. Second, and more importantly, the systemic risk argument does not require universal or even majority adoption of advanced AI to be a legitimate concern. Meng and Chen’s (2026) modelling suggests the systemic risk multiplier grows superlinearly as adoption increases, meaning the most dangerous phase may not be the current, relatively cautious stage described in the survey evidence, but a future stage where adoption crosses a threshold current data cannot yet observe. The reassuring survey evidence and the more concerning theoretical work are therefore not describing the same moment: one captures today’s baseline, the other what happens as that baseline shifts, so the two bodies of evidence are less a direct conflict than two points on the same trajectory.

Taken together, the discussion suggests a specific and fairly narrow conclusion. The current empirical footprint of AI-driven systemic risk in trading is real but still developing, and the theoretical mechanisms underpinning concern are well specified rather than speculative. The main open question is not whether these risks exist, but how quickly current institutional practice, which still leans conservative, will be overtaken by the more advanced, correlated and opaque systems the literature warns about.

Conclusion

In conclusion, through this extensive research project, we discovered that artificial intelligence is, and can prove to be, an incredible tool in the trading sector. These beneficial programs are more efficient than humans due to their lower reaction time and greater analytical capabilities, thus saving time, money and doubt on most day-to-day trades. Additionally, we found out that AI’s execution is also faster than humans due to no reaction speed necessities and its usually more accurate pattern recognition based on massive sets of data. Furthermore, artificial intelligence aids companies and traders through finding hidden patterns and drawing conclusions from large documents very quickly, allowing traders to trade more efficiently. However, these benefits have clear trade offs, namely herding, which causes unnecessary high market volatility and very rapid changes (Aldasoro et al., 2024). The other main disadvantage or potential hindrance is the inevitable need for data dependency; AI learns through the information it is given and through trial-and-error feedback loops, thus when a black swan event appears, the program is left confused as it has never encountered such activity before, rendering it potentially useless and harmful (Aldasoro et al., 2024). This reflects a broader concern around model opacity and data dependency that the literature identifies as central to AI’s systemic risk in financial markets (Aldasoro et al., 2024).

In response to the research question: “To what extent do AI-driven trading models improve trading decisions without increasing systemic risk?”, we conclude that to a small extent, it is fair. For AI to truly be beneficial for the trading industry, it must be regulated and used in addition to, or in coherence with, a trader, due to its high analytical skills and quick reactions. AI is a tool that must not be underestimated or used without proper training and knowledge of the extent to which it can impact trading decisions as a whole.

Aldasoro, I., Gambacorta, L., Korinek, A., Shreeti, V. and Stein, M. (2024) “Intelligent financial system: how AI is transforming finance”, BIS, Working Paper No. 1194. <https://www.bis.org/publications/working-paper-1194-intelligent-financial-system-how-ai-transforming-finance>

Cohen, G. (2022) “Algorithmic Trading and Financial Forecasting Using Advanced Artificial Intelligence Methodologies”, Mathematics, 10(18), p. 3302.

Cohen, S.N., Snow, D. and Szpruch, L. (2021) “Black-box model risk in finance”, arXiv: 2102.04757.

Danielsson, J., Macrae, R. and Uthemann, A. (2021) cited in: Maple, C., Szpruch, L., Epiphaniou, G., Staykova, K., Singh, S., Penwarden, W., Wen, Y., Wang, Z., Hariharan, J. and Avramovic, P. (2023) “The AI Revolution: Opportunities and Challenges for the Finance Sector”, arXiv: 2308.16538.

Danielsson, J. and Uthemann, A. (2023) “On the use of artificial intelligence in financial regulations and the impact on financial stability”, arXiv: 2310.11293 [preprint].

Elephant in the Boardroom (2025) “How AI driven data accuracy can improve efficiency and business growth”, Elephant in the Boardroom [online]. <https://www.elephantintheboardroom.com.au/blog/how-ai-driven-data-accuracy-can-improve-efficiency-and-business-growth>

Gong, S., Han, S., Kou, Z., Zhang, Y. and Zhu, L. (2026) “Artificial intelligence in financial markets: a literature review and research agenda”, China Accounting and Finance Review, 28(4), pp. 420-445.

Hansen, A.L., and Jung Lee, S. (2025). “Financial Stability Implications of Generative AI: Taming the Animal Spirits”, Finance and Economics Discussion Series 2025-090. <https://doi.org/10.17016/FEDS.2025.090>

Hilton, J. (2026) “Explainable AI in Algorithmic Trading: Why Black-Box Models Are a Regulatory Risk”, Medium [online]. <https://medium.com/@jessehilton612/explainable-ai-in-algorithmic-trading-why-black-box-models-are-a-regulatory-risk-c70a79dfafd7>

HSBC (n.d.) “Understanding market volatility”, HSBC UK [online]. <https://www.hsbc.co.uk/investments/understanding-market-volatility/>

Interactive Brokers (n.d.) “Toggle AI”, IBKR Campus [online]. <https://www.interactivebrokers.com/campus/contributors-categories/toggle-ai/>

Investopedia (n.d.) “Black swan”, Investopedia [online]. <https://www.investopedia.com/terms/b/blackswan.asp>

Li, L. (2021) “Financial trading with feature preprocessing and recurrent reinforcement learning”, arXiv: 2109.05283 [preprint].

Maple, C., Szpruch, L., Epiphaniou, G., Staykova, K., Singh, S., Penwarden, W., Wen, Y., Wang, Z., Hariharan, J. and Avramovic, P. (2023) “The AI Revolution: Opportunities and Challenges for the Finance Sector”, arXiv: 2308.16538.

Meng, S. and Chen, X. (2026) “Artificial Intelligence and Systemic Risk: A Unified Model of Performative Prediction, Algorithmic Herding, and Cognitive Dependency in Financial Markets”, arXiv: 2604.03272.

Paulise, L. (2026) “Is AI Replacing Finance Jobs? New Research and Career Trends”, Forbes [online]. <https://www.forbes.com/sites/lucianapaulise/article/will-ai-replace-finance-jobs/>

Qiu, Y. and Han, Q. (2026) “Representation Homogeneity and Systemic Instability in AI-Dominated Financial Markets: A Structural Approach”, arXiv: 2604.22818.

Sabkha, S. and Jbir, H. (2026) “Artificial Intelligence, Algorithmic Herding and Systemic Fragility in Financial Markets”, HAL Open Science, 05556083.

Sidley Austin LLP (2024) “Artificial Intelligence in Financial Markets: Systemic Risk and Market Abuse Concerns”, Sidley [online]. <https://www.sidley.com/en/insights/newsupdates/2024/12/artificial-intelligence-in-financial-markets-systemic-risk-and-market-abuse-concerns>

Svetlova, E. (2022) cited in: Maple, C., Szpruch, L., Epiphaniou, G., Staykova, K., Singh, S., Penwarden, W., Wen, Y., Wang, Z., Hariharan, J. and Avramovic, P. (2023) “The AI Revolution: Opportunities and Challenges for the Finance Sector”, arXiv: 2308.16538.

Tasdelen, I. (2026) “The Rise of AI-Driven Trading Systems”, Cubed [online]. <https://blog.cubed.run/the-rise-of-ai-driven-trading-systems-5a5e2abfa197>

Zhang, J. (2025) “Trading with the Devil: Risk and Return in Foundation Model Strategies”, arXiv: 2510.17165.