Academic interest in AI applied to finance has followed an explosive trajectory — from 6 articles in leading journals in 2019 to a peak of 35 in 2023, stabilising near 27 in 2024. ChatGPT (Nov 2022) and Meta's open LLaMA release (Feb 2023) triggered a fresh wave of transformer research in financial contexts. But volume came with a conceptual shift: from demonstrating ML's predictive superiority over linear benchmarks toward explaining the economic mechanisms underneath.
Gu, Kelly & Xiu (2020) set the baseline — ML roughly doubles the out-of-sample R² of linear models in cross-sectional return prediction. Kelly, Kuznetsov, Malamud & Xu (2025) mark the next step: embedding transformer architectures directly into stochastic-discount-factor construction, moving from prediction to economically-structured modelling. This review covers five domains and the cross-cutting tension that binds them.
A systematic read of 2019–2026.
The review spans 2019–2026 across Scopus, Web of Science and arXiv, anchored on a 138-paper FT50/UTD24 survey (2019–2024) and complemented by PRISMA reviews of forecasting, RL in portfolio management, NLP/LLMs, and AI-induced systemic risk. Selection favoured papers applying adaptive ML to a financial task with a substantive economic conclusion — not CS papers where financial data is a mere benchmark.
| Method class | Typical architectures | Share (FT50, 2019–2024) |
|---|---|---|
| Classical ML | SVM, Lasso, Ridge, kernels | 31 papers |
| Neural / deep learning | LSTM, CNN, Bi-LSTM, GNN | 25 papers |
| Reinforcement learning | DQN, PPO, Actor-Critic | 16 papers |
| Ensemble trees | Random Forest, XGBoost | 14 papers |
| NLP / transformers / LLM | BERT, FinBERT, GPT-4 | 14 papers |
Classical ML retains numerical dominance, but the direction is unambiguous — a steadily rising share for deep learning, RL and transformers in both academic and applied work.
From linear factors to nonlinear discounts.
Cross-sectional return forecasting is the single most active area — 46 of 138 papers. The consistent finding: ensemble trees and neural networks outperform linear five-factor models out-of-sample, evidence of significant nonlinearities and interaction effects among anomalies. The frontier then moves from prediction to theory-consistent ML — constructing stochastic discount factors directly (Kozak–Nagel–Santosh; Lettau–Pelger; Kelly et al. 2025) to ask which risk factors are genuinely priced, not merely which predict.
A parallel strand leverages alternative data — satellite imagery, transaction flows, geolocation. Traditional sources (Bloomberg, Compustat, CRSP) appear in 61 of 92 empirical papers; alternative data already in 45 — competitive pressure, since the value of public data is arbitraged away faster.
High out-of-sample R² rarely survives contact with reality: transaction costs and market impact erode academic figures into thin or negative trading alpha. Interpretability stays low — a strong predictor may be a risk premium, a behavioural bias, or a data-mining artefact.
From speed to intelligence.
Market microstructure and algorithmic trading account for 35 papers. Early work framed HFT through speed — latency arbitrage, optimal execution, spread tightening. The frontier has shifted to substantive intelligence: RL agents that learn market-making, liquidation, and dynamic hedging, adapting to regime change in real time. The dual market-quality effect is well established — tighter spreads and faster discovery on one side; short-term volatility spikes and flash-crash propagation on the other.
RLFrom execution to portfolio management.
- Optimal execution — minimise market impact when liquidating under a time constraint, with no assumption of linear or stationary impact.
- Market-making — post bid/ask dynamically, balancing inventory risk against spread income.
- Portfolio optimisation — DRL agents beat traditional strategies on risk-adjusted returns across S&P 500, CSI-300, NASDAQ.
- Derivative hedging — discrete hedging under transaction costs and fractional delta, inaccessible to continuous-time Black–Scholes.
Deployment is non-trivial: state normalisation drives generalisability (a crypto-robust model may degrade on equities), and hybrid RL + classical approaches (ARIMA, Kalman, stochastic control) consistently beat pure RL out-of-sample.
Credit risk, fairness & regulation.
Credit scoring is among the most practically mature subfields (27 papers). Random Forest and XGBoost beat logit on ROC AUC and Gini by capturing nonlinear interactions and non-standard data (transaction behaviour, digital footprint). Generative augmentation (CTGAN) addresses class imbalance — Recall 0.83 / AUC 0.91 on German Credit, with simultaneous fairness gains.
But the accuracy–fairness trade-off holds across datasets: better predictive quality does not systematically improve parity across protected attributes. Workable mitigations exist — post-hoc threshold optimisation (no retraining), SHAP-based explainability for GDPR / Basel right-to-explanation, and alternative-data models that cut decision variance for low-income clients sevenfold versus FICO. The EU AI Act (in force 2026) classes ML credit scoring as high-risk — mandatory registration, audit, explanation — shifting preference toward interpretable architectures.
From dictionaries to transformers.
NLP is the fastest-expanding toolset — from Loughran–McDonald lexicons through FinBERT to GPT-4 and fine-tuned domain models. Domain-specialised models beat general ones on financial sentiment; the qualitative jump came with GPT-4: Kim et al. (2024) show LLMs — without industry context — outperform human sell-side analysts at predicting the direction of EPS changes, with GPT-based strategies generating significant alpha.
Principal applications, and the one hard limit:
- Financial-statement analysis — extract insight from structured disclosures, even under non-standard formats.
- Earnings automation — fine-tuned LLMs with RAG approach GPT-3.5 on research-note generation.
- Knowledge distillation — compact task-specific models resolve latency and cost without material accuracy loss.
- Hallucination — LLMs invent numbers and facts when parsing filings; verification, retrieval grounding and uncertainty quantification mitigate but do not eliminate it.
The defensible claim is not “LLMs predict prices.” It is that they encode and route semantic shocks into a quantitative pipeline.
New instability, new monitoring.
Homogeneous AI strategies create qualitatively new systemic channels. Algorithmic herding arises from model convergence — similar data and architectures produce correlated signals and simultaneous order flow (the 2010 Flash Crash, 2015 China, 2021 GameStop, 2022 crypto). Algorithmic collusion — Calvano et al. (2020) show RL agents can sustain supracompetitive spreads with no explicit coordination, a fundamental problem for competition policy.
On the monitoring side, Graph Neural Networks capture network topology and temporal dynamics together: an LSTM + GNN early-warning framework reached 92.8% alert accuracy and 89.5% contagion-path precision, ~34% better tail sensitivity than VaR. RegTech/SupTech extends this to AML/CFT on transaction graphs, fair-lending auditing, and real-time surveillance of spoofing, wash trading and layering — within a three-lines-of-defence governance frame with human-in-the-loop, dynamic pre-trade limits and kill-switches.
The central unresolved tension.
Today's XAI is mostly post-hoc — maximise performance, then explain with SHAP, LIME or attention. That does not guarantee economic meaning: high SHAP importance can reflect a genuine factor or a spurious correlation. The more promising path is inherently interpretable architectures (GA²M, Neural Additive Models), still niche in finance. Regulation pushes the same way — Basel IV documented model logic, GDPR and the EU AI Act explainability standards — creating incentives for natively interpretable design over post-hoc patches.
What to build next.
Structurally grounded ML. Source, not just R².
Identify whether an ML-discovered anomaly is a risk premium, mispricing, or friction — with structural restrictions, not model-class horse races.
Safe RL & multi-agent equilibrium. Constraints in the loop.
RL that guarantees position/VaR/leverage compliance during training; theory for when heterogeneous AI agents collude or destabilise, and how to design circuit breakers and diversity requirements.
Federated learning & financial foundation models. Privacy and scale.
Train credit/risk models without centralising sensitive data (raw data stays inside the perimeter); pretrained models for heterogeneous financial time series, robust to nonstationarity and adversarial dynamics.
Key takeaways.
- ML roughly doubles the out-of-sample R² of linear asset-pricing models — but the gain often dissolves under transaction costs and market impact.
- The frontier shifted from prediction to theory-consistent ML — embedding learning into stochastic-discount-factor construction.
- RL reshaped execution, market-making and hedging; hybrid RL + classical beats pure RL out-of-sample.
- LLMs can rival sell-side analysts at EPS-direction forecasting, but hallucination makes verification and retrieval grounding non-optional.
- GNNs deliver a new tier of systemic-risk early warning; AI also creates new risk via herding and learned collusion.
- The defining unresolved tension is predictive power versus economic interpretability — and regulation now prices it in.
Technical progress outpaces both theory and the regulatory frameworks meant to govern it. The most productive strategy for 2026–2030 is not another demonstration of predictive superiority — it is bridging ML technique and economic mechanism: disciplined welfare analysis, structural identification, and an evidentiary basis for sound regulatory design.