What is AI in share trading?
An encyclopaedic overview: what people mean when they say “AI trading” in 2026, the two distinct technology families under that label, what each does well, and where each fails.
“AI trading” is a marketing label draped over at least three distinct technologies. Some of it is genuinely new. Some of it is decades-old statistical arbitrage relabelled. Some of it is a large language model reading news. Distinguishing the three is the first step to reasoning clearly about what a share-trading system can and cannot do.
Section ITwo families under one label
In the 2026 market the phrase “AI in share trading” is applied to two technology families that share almost no components:
- Large language models (LLMs) and LLM-based agents. Generative models trained on text (and increasingly on charts, tables and audio). They read filings, digest news, tag sentiment, extract structured facts, argue for and against a thesis, and coordinate workflows.
- Machine learning applied to numerical market data plus classical quantitative methods. Statistical arbitrage, factor models, tree-ensembles on returns and volume, deep learning on order-book microstructure, reinforcement learning for execution. The technologies used by systematic hedge funds for the last twenty-plus years, with continuous refinement.
The confusion between the two matters because the risks, costs, latency profiles and regulatory treatment differ substantially. LLM output is textual and probabilistic; ML signals are numerical and (in principle) reproducible. A quant strategy can run at sub-millisecond latency; an LLM cannot, and is not appropriate for latency-sensitive execution paths. The correct pattern is usually to use each where it belongs.
Section IILarge language models
The public benchmark literature has moved past “can an LLM read a 10-K?” and now measures where they break. FinanceBench (Patronus AI) — a 150-sample gold set of real 10-K questions — found GPT-4-Turbo, in closed-book mode, correct on 9% of questions with 88% refusals; with a naive vector-store RAG configuration, 50% correct; with the whole document in a long context window, 79%; with an oracle chunk, 85%. Patronus’s summary is stark: “GPT-4-Turbo with retrieval incorrectly answered or refused to answer 81% of questions.”
Two 2025 papers deserve to be cited every time someone claims an LLM “understands” filings. FailSafeQA (arXiv 2502.06329) found OpenAI o3-mini fabricated information in 41% of tested cases even with retrieval, and the finance-specialised Palmyra-Fin-128k-Instruct failed robustness in 17%. FAITH (arXiv 2508.05201) shows top models collapse “from 95.6% accuracy on simple lookups to near 0% on multivariate calculations” on intrinsic tabular hallucination tests.
For sentiment and lightweight classification, the older specialists remain competitive. FinBERT still hits roughly 80% on Financial PhraseBank with only 250 fine-tuning examples. FinGPT reports F1 0.7488 on the 2025 sentiment evaluation, matching or beating IBM Granite 3.0 and Llama 3.1. BloombergGPT holds 75.07 on FiQA sentiment.
The one-line summary
LLMs are strong at extraction and classification over financial text. They are weak at numerical multi-step reasoning. Do not let one do arithmetic without a calculator tool.
Section IIIMachine learning & classical quant
The methods most systematic funds actually use to make money are older and more numerical: linear factor models (Fama-French and successors), gradient-boosted trees on engineered features, LSTM and transformer models on price and volume series, reinforcement learning for execution, deep learning on limit-order-book microstructure. These are not “AI” in the ChatGPT-era sense but they are the working horses of the industry. They dominate high-frequency and mid-frequency strategies because they are numerical, deterministic given a random seed, and cheap enough per prediction to run at market speeds.
Classical quant strategies — statistical arbitrage, pairs trading, mean-reversion, momentum, cross-sectional factor tilting — do not require deep learning at all. Many of the largest systematic firms still generate most of their returns from linear models over carefully engineered features, with more sophisticated architectures used at the margin. A retail builder should be careful not to reach for a neural network when a rolling z-score does the job.
Section IVThe retail-to-institutional spectrum
Where on the retail-to-institutional spectrum a builder is operating determines almost every design choice.
- Retail builder, one account. End-of-day or swing trading, paper execution first, human approval gate on every order. Free or cheap market data (delayed for research, Alpaca IEX for live paper). Any modern broker API works. Regulatory perimeter: personal use.
- Small professional, managing one’s own capital plus friends and family. Regulatory questions begin the moment someone else’s capital is involved. See Regulatory boundary.
- Signal-selling or subscription service. The financial-promotions regime bites. In the UK the FCA’s Consumer Duty and the Marketing Rule (US) apply.
- Discretionary or non-discretionary investment management. Full regulatory authorisation (FCA Part 4A, SEC RIA registration, or equivalent) becomes mandatory.
- Institutional systematic fund. Colocation, direct market access, low-latency infrastructure, teams of quants. Outside the scope of this reference.
Section VBenchmarks and honest limits
Public multi-agent LLM trading papers report headline Sharpe ratios that do not survive scrutiny. TradingAgents (arXiv 2412.20138) reports Sharpe roughly 8.2 on Apple in Q1 2024 versus 2.05 for the best baseline. The window is a three-month bull tape on a mega-cap tech name; transaction costs are not modelled; GPT-4’s training cutoff post-dates Q1 2024 so the sentiment analyst is reading news whose outcome is already in the model weights.
The 2025-2026 meta-critique is beginning to arrive. LiveTradeBench (arXiv 2511.03628) and the ACM AIFT 2026 reproducibility paper put TradingAgents on GOOG May-July 2025 at 15-18% versus a 19% buy-and-hold. Reproducibility scored 1 out of 5 for TradingAgents / FinVision, 0 for FinMem, 2 for FinCon and FinAgent. Profit Mirage (arXiv 2510.07920) shows reported alpha collapses on truly post-cutoff data. A recent survey (arXiv 2603.27539) finds only 1 of 19 published studies modelled transaction costs at all.
Adopt the architecture from the published multi-agent literature (analyst-debate-trader-risk-manager). Do not adopt the P&L numbers.