Something for the Weekend – The Quiet Revolution

How Agentic Quant Research, Multi-Agent HFT and Reinforcement-Learnt Alpha Weighting are reshaping how markets trade: AI in Trading – 20th September 2025

Rebecca Healey
September 20, 2025

While political headlines focus on US-UK trade pacts and big-ticket investments, a quieter but more profound shift is already underway inside trading firms. AI is no longer a side experiment; it is transforming how markets trade. From multi-agent large-language-model (LLM) frameworks for high-frequency trading to reinforcement learning that re-weights trading signals in real time, here are the 5 key developments to watch this week:

1. QuantAgent: Multi-Agent LLMs Move into High-Frequency Trading

ArXiv:2509.09995 introduces QuantAgent, a multi-agent LLM framework built specifically for high-frequency trading (HFT). Four specialised agents—Indicator, Pattern, Trend and Risk—each analyse a different facet of short-term market dynamics before voting on live trade decisions in instruments such as Nasdaq futures and Bitcoin.

Unlike earlier LLM systems (for example TradingAgent or FINMEM) that focus on long-term fundamental reasoning, QuantAgent targets short-horizon, precision-critical trading, where structured signals such as technical indicators, chart patterns and micro-trends matter more than macro sentiment. In zero-shot tests across ten instruments it delivered higher predictive accuracy and cumulative returns over four-hour intervals than random baselines.

Why this matters for Trading:
This moves AI beyond medium-term forecasting into sub-second execution. A built-in risk agent shows how AI HFT can self-monitor exposures in real time.  This raises the bar for latency-sensitive trading desks and shifts the competitive edge from raw speed further towards smarter pre-trade decision-making.

2. Multimodal Reinforcement-Learnt Asset Allocation

With roughly 90 % of global trading volume now algorithmic, high-frequency AI agents are only one side of the equation. New research (AINvest) highlights multimodal LLMs fused with reinforcement learning to improve dynamic asset allocation.

These systems blend structured market data with unstructured inputs such as news sentiment and macro forecasts, continuously adjusting how capital is distributed across asset classes. Whereas QuantAgent optimises how to trade at very high speed, these RL platforms optimise where to allocate capital over longer horizons – guiding investment strategy rather than micro-second execution. Deep-learning models also mine unstructured data – earnings calls, central-bank statements – to anticipate price moves within milliseconds, making markets simultaneously more efficient and more volatile.

Why this matters for Trading:
This approach closes the gap between static mean-variance models and fully dynamic allocation, enabling trading desks to play a greater role in responding intraday to regime shifts in macro or sentiment.  But this will also demand forensic real-time data governance – garbage in still means garbage out.

3. PPO – LLM-generated alpha on the fly

While asset-allocation RL decides where to invest, adaptive alpha weighting decides how to combine and size individual trading signals inside a portfolio. arXiv:2509.01393 introduces a reinforcement-learning “meta-strategy” that uses Proximal Policy Optimisation (PPO) to re-weight LLM-generated alphas on the fly.

Using deepseek-r1-distill-llama-70b model in live equity tests on Apple, HSBC, Pepsi, Toyota, and Tencent, PPO was used to adjust weights in real time. The results outperformed equal-weight and benchmark portfolios, proving that the real challenge in quantitative trading is shifting from merely finding signals to continually reallocating capital as market regimes evolve. PPO provides a self-adjusting mechanism that reduces drawdowns when individual alphas lose their edge, making it ideal for large quant funds managing hundreds of concurrent signals.

At the same time, new technical models such as Trading-R1 (arXiv:2509.11420) and DeltaHedge (arXiv:2509.12753) push AI-driven trading further: Trading-R1 blends LLM reasoning with reinforcement learning to deliver risk-aware, evidence-grounded trading decisions, while DeltaHedge offers a multi-agent framework for options hedging and portfolio optimisation, delivering stronger performance in volatile conditions.

Why This Matters for Trading:

• Real-time adaptability: Reinforcement learning enables models to re-weight signals on the fly, making strategies more resilient to sudden changes in conditions.
• Integrated risk controls: Trading-R1 and DeltaHedge embed risk awareness and hedging inside the AI itself, moving beyond black-box prediction to interpretable, evidence-based decision systems.
• Scalability for large quant funds: These approaches let firms running hundreds of signals or complex options books automate tasks once done manually, improving efficiency and consistency.
• Setting a higher bar for oversight and governance: Faster idea generation and automated execution heighten the need for robust validation and regulatory compliance, reshaping how quant shops design and monitor strategies.

 4. Crypto Desks and Fixed-Income Catch-Up

More than half of institutional crypto managers now deploy AI-driven automated strategies, combining order-book analytics and sentiment detection to run 24/7 strategies (https://www.ainvest.com/news/ai-redefines-crypto-trading-algorithms-outpace-human-instinct-2025-2509/?utm) . Even the traditionally conservative fixed-income markets are now catching up. As Dan Barnes reports, bond traders foresee a data-science and AI boom in 2026, promising to turn a fragmented, opaque market into one where pricing, execution and risk management can be automated. Key benefits include deeper liquidity discovery, real-time risk control, bespoke and more accurate pricing, smarter venue and order routing, and lower trading costs through continual learning from past executions.

Why This Matters for Trading:

· Deeper liquidity discovery: AI can aggregate and analyse dealer quotes, trade prints and market news to pinpoint the most executable prices across many venues.

· Real-time risk control: It can instantly detect sudden rate moves or market shocks and trigger automatic hedging or portfolio rebalancing.

· Faster, more tailored pricing: By modelling yield-curve shifts, credit spreads and macro data in real time, AI can deliver customised prices on demand, replacing generic quotes that often miss market nuances.

· Smarter venue and order routing: Autonomous agents can locate liquidity, negotiate RFQs and split orders efficiently across all-to-all and 24/7 platforms.

· Lower costs and improved execution: Learning from historical trade outcomes, AI reduces market impact and continually optimises fill quality.

 5. We’ve written about this before but it appears worth repeating. The AI winter may be approaching – recent data from the US Census Bureau’s Business Trends and Outlook Survey shows that large firms (250+ employees) are beginning to slow down AI adoption (https://www.apolloacademy.com/ai-adoption-rate-trending-down-for-large-companies/) As Torsten Sløk highlights after a year of experimentation, businesses may be pausing to reassess how and where AI truly adds value – but in trading it is recognising that AI represents the next stage of automation: AI is now designing, selecting and executing trades across both traditional and digital assets. Yet for every positive announcement of AI model capability, there are the twin challenges of governance and oversight (ensuring human supervision and compliance) and regime risk (perfectly executed trades can still be wrong if market conditions change suddenly yet again). Cue more from the FCA in how they plan to adapt regulation to meet these changes – read more here – https://www.fca.org.uk/news/speeches/regulating-growth-future-now.

Trading investment strategies is no longer about who has the largest model or who can get to the market faster; it is about smarter speed and the ability to adapt to changing dynamics – fusing AI, risk controls and market microstructure into a sustainable trading edge.  Just a few years ago, the market was claiming that AI would never be involved in direct execution – we can see just how much this has changed.  In the FIX AI Working Group we are continuing to look at what standards will be needed in the future to manage this – such as tagging algo versions as well as authentication tokens to ensure firms can track any algo changes, mandate roll-backs when necessary and maintain secure, client-specific data access. If you’d like to join the group, please let me know.

As always, thank you for reading, and let me know what you found most useful, what you disagreed with, and what you would like to see more of next time.

Best wishes

Rebecca