Dr Jinghan Meng

11 April 2026

If, in the future, the AI agents of various funds in Central, Hong Kong can finish reading central bank statements, earnings call transcripts, and global news within the same second, and make similar directional trading decisions accordingly, will the market become more efficient or more vulnerable? This is no science-fiction fantasy but a reality fast approaching the financial industry.

At this stage, generative artificial intelligence (GenAI) is mainly used in summarizing study reports, collating information, writing codes, and analysing texts. However, when these capabilities are further integrated with memory, reasoning, planning, tool utilization, and continuous execution modules, and embedded into investment research, risk management, and trading processes, financial institutions will move from “signal generation” to more autonomous agentic AI systems. (see Notes 1, 6, and 8)

“Alpha anxiety” in the age of AI

This change has given rise to the phenomenon of “alpha anxiety”. Alpha refers to excess returns. The anxiety is not so much about whether AI is smarter than people, but about the fact that AI is rapidly reducing the cost of replicating some investment research tasks. The tasks of organizing public information, interpreting texts, tracking holdings, and identifying styles, which used to take research teams a long time to complete, can now increasingly be automated by models (see Note 4). Some studies suggest that, using only publicly disclosed holdings and macroeconomic data, AI can already mimic a considerable proportion of top asset managers’ trading behaviour. (see Note 2)

As more institutions can extract similar signals from similar data, the previously scarce informational edge becomes more susceptible to competitive erosion, and alpha also becomes harder to sustain. AI may not necessarily put an end to alpha but it is indeed shortening alpha’s half-life, forcing institutions to invest in more computing power, more expensive data, and more complex architectures just to maintain existing returns. This is a classic example of Red Queen competition.

Algorithmic convergence and systemic fragility

A greater cause for concern than a single model committing errors is a large number of models simultaneously “getting the same thing right”. When different institutions rely on similar foundation models, similar news sources, the same market data, similar risk constraints, and similar optimization objectives, they may appear to be competing with one another, but in moments of stress, may converge towards highly similar trading responses. On the one hand, GenAI reduces information asymmetry by rapidly transforming unstructured information previously scattered across text, speech, and narratives into tradable signals. On the other hand, it may also enable market participants to arrive at more similar judgments in a shorter time, thus increasing the risk of strategy convergence and model homogenization. (see Notes 1, 6, and 7)

Under normal market conditions, such technological advancement can help to expedite price discovery. However, once markets face pressure, their procyclicality could also intensify. If more and more model-driven traders interpret central bank language, earnings guidance, and macroeconomic data using similar logic, and synchronously adjust positions under similar stop-loss rules, margin requirements, and value-at-risk limits, market liquidity could contract simultaneously within a short time. Existing research shows that while algorithmic trading, in normal circumstances, can accelerate price discovery and the incorporation of information into prices, under stress scenarios, it could exacerbate the vulnerability of liquidity. If different strategies rely more heavily on similar signals and similar execution rules, the improvement in market efficiency may come at the cost of a more fragile market microstructure. This is exactly a paradox of AI finance: rational optimization at the micro level may not necessarily lead to stability at the macro level. (see Notes 3 and 7)

From chasing alpha to pursuing “responsible alpha”

That is exactly why I propose the concept of “responsible alpha”. In the future, valuable alpha should not just outperform the market in the short term but should also be kept within boundaries that are explainable, auditable, and open to intervention, without unduly amplifying systemic risk. In other words, instead of treating risk control as an add-on, “responsible alpha” internalizes governability as part of alpha. With the increasing commoditization of signal extraction, the truly scarce capability will not simply be building ever more opaque black boxes, but demonstrating why one’s AI is trustworthy: what data it uses, what workflow it follows, who can review it, and who can put a stop to it if anything goes wrong. When alpha comes to resemble a replicable public technology, governance capability, auditing capability, and human intervention capability, it may instead become a new private moat. (see Note 5)

Hong Kong’s policy approach in the past couple of years has provided the institutional groundwork for such “responsible alpha”. In August 2024, the Hong Kong Monetary Authority (HKMA), in conjunction with Cyberport, launched the GenA.I. Sandbox, clearly setting out a risk-based approach and emphasizing that high-risk decision-making must retain a human-in-the-loop model. In August 2025, the Bank for International Settlements (BIS) Innovation Hub, Hong Kong Centre, the HKMA, and the UK Financial Conduct Authority launched Project Noor, focusing on the application of GenAI and advanced algorithmic models in the financial system to address the AI explainability problem. This means that the regulatory approach no longer merely requires institutions to explain AI, but is beginning to enhance supervisors’ own ability to understand black boxes. In March 2026, Hong Kong upgraded the sandbox to GenA.I. Sandbox++, extending coverage to securities, asset management, insurance, MPF, and other areas. As a result, the sandbox is no longer a testing ground but also an institutional project through which regulators and the market jointly define safety parameters.

Of course, the launch of the sandbox cannot simply be a showcase for innovation but is meant to be a testing ground for governance capability. In terms of agentic AI and quantitative models, what is really being tested is not model accuracy but the entire risk management chain: whether the data is traceable, whether version updates leave an audit trail, whether orders will be cancelled simultaneously under stress scenarios, whether model drift can be detected in time, and whether human intervention or emergency shutdown is possible when necessary. Future competition in the asset management industry will hinge on institutions’ ability to incorporate model risk into their governance frameworks, rather than on model predictive capability alone.

Aiming to enhance the trustworthiness of alpha

The next round of financial competition will be about far more than who adopts AI first; it will be about who can prove sooner that their AI is reliable when making money and controllable when it fails. A World Economic Forum white paper released in 2025 notes that many institutions are in fact still in the transitional stage from experimentation to scaled implementation. What truly determines success or failure is often not the model itself, but whether trust, self-governance, talent, cybersecurity, and digital infrastructure are in place. (see Note 8) From this perspective, Hong Kong’s potential advantage lies not only in deploying AI earlier, but in institutionalizing model governance, stress testing, audit trails, and human intervention mechanisms earlier, gradually turning them into a common language across markets, regulators, and institutions.

In the final analysis, if markets ultimately regard transparency, accountability, and governance maturity as attributes worth paying for, then “trust” itself may become the most important intangible asset of the next generation of international financial centres. What Hong Kong should pursue is not higher alpha, but alpha that global capital can trust.

Note 1：Aldridge, I., An, J., Burke, R., Cao, M., Chien, C. Y., Deng, K., … & Zheng, W. (2025). Agentic artificial intelligence in finance: A comprehensive survey [Working paper].

Note 2：Cohen, L., Lu, Y., & Nguyen, Q. H. (2026). Mimicking finance (NBER Working Paper No. 34849). National Bureau of Economic Research.

Note 3：Dou, W. W., Goldstein, I., & Ji, Y. (2025). AI-powered trading, algorithmic collusion, and price efficiency (NBER Working Paper No. 34054). National Bureau of Economic Research.

Note 4：Gu, S., Kelly, B., & Xiu, D. (2020). Empirical asset pricing via machine learning. The Review of Financial Studies, 33(5), 2223–2273.

Note 5：Fabozzi, F. A., & López de Prado, M. (2025). Implementing AI Foundation Models in Asset Management: A Practical Guide. Journal of Portfolio Management, 52(2).

Note 6：Mo, H., & Ouyang, S. (2025). (Generative) AI in financial economics. Journal of Chinese Economic and Business Studies, 23(4), 509–587.

Note 7：International Monetary Fund. (2024). Global financial stability report. https://impact.hkubs.hku.hk/wp-content/uploads/2026/05/291953-textrevised.pdf

Note 8：World Economic Forum. (2025). AI in action: Beyond experimentation to transform industry. AI Governance Alliance, in collaboration with Accenture.