Quantitative analyst

Euro area, banks supervised by the ECB that use internal models. The guide expects the main stakeholders working with machine-learning-based internal models to have sufficient skills, capabilities and expertise in the techniques to manage model risk; internal validation to effectively challenge the modelling decisions taken on those models, assessing whether their complexity is justified; and firms to establish monitoring that prevents the automatic implementation of material changes. It states supervisory expectations ('should') for bank capital models, not trading research.

Failure, rollback, regulation or cost is suppressing adoption. Can lower an assessment or widen its uncertainty.

A benchmark in which models turn strategy descriptions into reproducible backtests, evaluated on 23 language models. The paper reports that the top model reached an overall accuracy of 67.41%, and that while simpler metrics such as win rate see higher accuracy, complex statistical indicators such as volatility and Sharpe ratio remain 'disaster zones' across all models. It is a fixed benchmark, not production work.

A demo, benchmark or paper shows the task can be done. Updates what the technology can do — not what employers will do.

A benchmark of 17 models across 12 strategies that turns strategy specifications into trading-system code and iterates on it. The abstract reports that top models achieve validity above 91.7 percent with strong aggregate scores, that iteration also induces code convergence, and that human oversight remains essential for critical strategies requiring solution diversity and ensemble robustness. It is a benchmark, not production work.

A demo, benchmark or paper shows the task can be done. Updates what the technology can do — not what employers will do.

Man Group, a systematic asset manager. The article says its AlphaGPT system writes production-grade Python code leveraging the firm's research tools and can interact with proprietary databases, that to date it has produced signals that meet the firm's standards and pass the same evaluation thresholds required for human-generated research, and that AI-generated signals undergo dual-track validation. It gives no count of signals and is the firm describing its own tool to investors.

An employer has put it into production. Can move the baseline — weighted by scale and how similar the setting is.

A research framework in which agents propose, implement and test factors and models. The abstract reports that it achieves up to 2X higher annualised returns than classical factor libraries using 70% fewer factors, and outperforms deep time-series models on real markets. The results are backtests, not live trading, and the authors' employer distributes the framework.

A demo, benchmark or paper shows the task can be done. Updates what the technology can do — not what employers will do.

United Kingdom. Based on conversations with market participants, the committee's review of principal trading firms and systematic hedge funds says that currently, where such techniques are deployed, they are implemented largely as rules-based systems with human-in-the-loop, and warns that the reasons for positions may not be well understood by human risk managers at the firm. It gives no figures.

An employer has put it into production. Can move the baseline — weighted by scale and how similar the setting is.

United Kingdom, banks with approval to use internal models. The statement says the validation function should operate independently from the model development process and from model owners, and that validation should be performed by staff with the requisite technical expertise and sufficient familiarity with the line of business using the model. The policy took effect on 17 May 2024. It is written as supervisory expectations ('should'), and it does not measure validation work.

Failure, rollback, regulation or cost is suppressing adoption. Can lower an assessment or widen its uncertainty.

The Netherlands, four major proprietary trading firms. The regulator says its study found that 80% to 100% of their trading algorithms aimed at liquid and standardised instruments, such as futures and equity, rely on machine learning models, and that it expects firms to be able to explain why orders were sent to the market; its report adds that reinforcement-learning trading algorithms are not used in practice yet. The figures are what four firms reported about classic machine learning, not large language models.

An employer has put it into production. Can move the baseline — weighted by scale and how similar the setting is.

European Union, investment firms engaged in algorithmic trading. Article 5 says a person designated by the senior management of the investment firm shall authorise the deployment or substantial update of an algorithmic trading system, trading algorithm or strategy, and that records must allow the firm to determine the person who made each change and the person who approved it; Article 3 requires risk and compliance staff with sufficient authority to challenge staff responsible for algorithmic trading. It names who authorises; it does not govern research methods.

Failure, rollback, regulation or cost is suppressing adoption. Can lower an assessment or widen its uncertainty.

United States, FINRA member broker-dealers. The rule says each associated person who is primarily responsible for the design, development or significant modification of an algorithmic trading strategy relating to equity, preferred or convertible debt securities, or responsible for the day-to-day supervision of such activities, shall be required to register as a Securities Trader; the regulatory notice adds that if a firm directs a third party to significantly modify a strategy, that direction must also be by a Securities Trader. It names who answers for a strategy; it does not govern how strategies are researched.

Failure, rollback, regulation or cost is suppressing adoption. Can lower an assessment or widen its uncertainty.

If you are starting out, expect AI to write much of your research and backtest code and to propose signals, and expect your value to lie in knowing when the numbers are wrong — the risk metrics models still miscompute — and in the validation and accountability roles the rules reserve for qualified people.

Expect to review and approve far more machine-generated research than you could write, and to answer for it. Firms using AI in research describe putting its output through the same tests as human work; designing those tests is where experience counts.

Stay and strengthen

Stay, and own the validation of machine-generated research

Firms put AI-generated signals through the same thresholds as human research; someone has to set those thresholds and catch what passes them wrongly.

Reshape the role

Move towards model risk and validation

US, EU and UK supervisors require independent validation by qualified people, and now write expectations specifically for machine-learning models.

Adjacent move

Move into algorithmic trading oversight

Rules require a designated person to authorise algorithm deployments and staff with the authority to challenge algorithmic trading.

Not on present evidence. AI now writes research code and proposes signals, and one asset manager says its AI-generated signals pass the same thresholds as human research. But benchmarks show models failing on basic risk metrics, and rules in the US, EU and UK require named, qualified people to design, authorise and independently validate trading and risk models.

We do not answer that with a number of years. A signal to watch instead: whether AI-generated strategies start going live without a designated person authorising them, and whether supervisors accept machine validation in place of independent people. Today the rules recorded here require both.

Parts of it. Man Group says its AlphaGPT writes production-grade code and has produced signals that pass its human-research thresholds, and a research agent reached up to twice the annualised returns of classical factor libraries in backtests. On BacktestBench, though, the best model reached 67.41% overall accuracy and all models struggled with volatility and Sharpe ratio.

The evidence recorded here points to a changing job rather than a vanishing one: more machine-generated research to test, and validation and accountability roles that US, EU and UK rules keep with qualified people. No official projection separates quants from other financial analysts.