Quantitative analyst — tasks, one by one

Signal research and strategy design

AI now proposes research as well as running it. Man Group says its AlphaGPT system has produced signals that meet its standards and pass the same evaluation thresholds required for human-generated research, with AI-generated signals put through dual-track validation. Machine learning itself is not new here: four large proprietary traders told the Dutch markets regulator that 80% to 100% of their algorithms for liquid instruments rely on it. In a research benchmark, an agent reached up to twice the annualised returns of classical factor libraries in backtests.

The firm's account is its own marketing to investors; the regulator's figures are what four firms reported; the benchmark is a backtest, not live trading.

Backtesting and quantitative code

Models write much of the code but miss on the numbers that matter. Man Group says its system writes production-grade Python using the firm's research tools. On BacktestBench, the best of 23 models reached 67.41% overall accuracy, and volatility and Sharpe ratio remained 'disaster zones' across all models; on another benchmark, top models produced valid strategy code more than 91.7% of the time while the authors concluded that human oversight remains essential.

Benchmarks with fixed tasks and one firm's account; nothing here measures how much code quants now write themselves.

Pricing and risk models

Supervisors now expect machine learning inside these models and set conditions on it: the European Central Bank's guide to internal models expects the people working with machine-learning models to have sufficient expertise in the techniques, and firms to prevent material changes from being implemented automatically. A central bank review found that where AI techniques are deployed in trading, they are largely rules-based systems with a human in the loop.

Supervisory expectations and a review based on conversations with firms; neither measures how models are built.

Model validation and challenge

Supervisors put validation on independent, qualified people. UK expectations say the validation function should operate independently from model development and be staffed by people with the requisite technical expertise; the ECB expects internal validation to effectively challenge the modelling decisions taken on machine-learning models, including whether their complexity is justified.

Supervisory expectations written with 'should', for banks with internal models; they do not measure validation work.

Deploying strategies and answering for them

The rules name the person. In the US, anyone primarily responsible for designing, developing or significantly modifying an algorithmic trading strategy must register as a Securities Trader; in the EU, a person designated by senior management must authorise the deployment or substantial update of a trading algorithm, and records must show who made and who approved each change.

Rules for US broker-dealers and EU investment firms; they do not measure how strategies are developed.