AI in finance is fractured across forecasting, portfolio design and ESG analysis, but the next advance will come from integrated decision architectures that link signal extraction to allocation and sustainability objectives.

Main Finding

AI in finance organizes around two complementary decision lines—short-horizon market prediction and slower-moving firm-level feature inference (e.g., financing constraints, ESG). Rather than treating forecasting, portfolio allocation, and ESG/corporate-finance analysis as separate literatures, the review argues they should be integrated into a decision architecture where AI extracts multimodal signals, converts them into constrained allocation rules, and evaluates outcomes under economic objectives (returns, risk, ESG). Gains from advanced models (LSTM, transformers, ensembles) are conditional: improved statistical fit does not guarantee better economic decisions once costs, regime shifts, and optimization sensitivity are accounted for.

Key Points

• Two foundational AI roles in finance:
  • Market prediction: exploitable patterns in prices, returns, volatility, sentiment, cross-sectionals for tradable signals and short decision cycles.
  • Firm-level feature identification: infer latent corporate attributes (financing constraints, governance, ESG) that inform capital allocation and cost of capital.
• Evolution of forecasting methods:
  • Classical models (ARIMA, GARCH) remain important baselines due to interpretability and low estimation variance.
  • Machine learning (SVM, trees, XGBoost, random forests) helped relax linearity assumptions.
  • Sequence models (RNN, LSTM) capture temporal dependencies; transformers and Informer variants can model long-range dependence and multimodal inputs but are data- and compute-intensive.
  • Hybrid and ensemble architectures combine prediction modules with optimization/selection layers.
• Multimodal inputs matter: textual disclosures, earnings calls, news, and social media add information beyond market histories—important for bankruptcy, ESG, and firm-heterogeneity tasks.
• From prediction to allocation:
  • Forecasts must be translated into allocation under transaction costs, turnover limits, risk budgets, liquidity, and governance constraints.
  • Common pipelines: forecast → preselection/ranking → optimizer (mean–variance, omega, risk-parity, hierarchical allocation) or end-to-end RL/online learning for dynamic rebalancing.
  • Risk estimation (Ledoit–Wolf shrinkage, factor models) often drives realized performance more than marginal improvements in return forecasts.
• Evaluation and target-definition:
  • Statistical metrics (RMSE, directional accuracy) are insufficient; economic evaluation (Sharpe, realized returns after costs, turnover, drawdown) is necessary.
  • Choice of target (one-step return, volatility, tail risk, ESG-score change, distress prob.) materially alters model design and usefulness.
• Major methodological risks:
  • Overfitting, sample dependence, regime instability.
  • Interpretability deficits and accountability when high-stakes financial/ESG decisions are automated.
  • Institutional and data-quality dependence (rating heterogeneity, measurement error in ESG).
• Central recommendation: build integrated frameworks that jointly model market prediction, portfolio design, and firm-level sustainable-finance analysis under explicit assumptions about data quality, decision frequency, and accountability.

Data & Methods

• Data types discussed:
  • Market data: prices, returns, volumes, microstructure.
  • Structured firm data: financial statements, ratios, balance-sheet variables.
  • ESG scores and sustainability indicators; note rating divergence and measurement heterogeneity.
  • Unstructured text: news, earnings calls, annual reports, social media.
  • Alternative data (implied, e.g., sentiment indices; not deeply enumerated).
• Forecasting/modeling methods:
  • Classical time-series: ARIMA, autoregressive models, volatility models (GARCH).
  • Supervised ML: SVM, decision trees, random forests, XGBoost.
  • Deep learning: feedforward nets, stacked autoencoders.
  • Sequence models: RNN, LSTM (gated memory for lag structures).
  • Attention-based models: Transformers, Informer (for long-range and multimodal sequences).
  • Hybrid/ensemble pipelines: prediction modules feeding mean–variance, omega, or other optimizers; preselection/ranking layers to control turnover.
  • Risk estimation: Ledoit–Wolf shrinkage, latent factor models.
  • Decision-oriented learning: reinforcement learning, online learning for direct policy optimization and dynamic rebalancing.
• Evaluation approaches:
  • Statistical losses (RMSE, classification accuracy) and economic metrics (Sharpe, drawdown, turnover-adjusted returns).
  • Stress/regime testing, out-of-sample and transaction-cost-aware backtests emphasized.
• Anchor studies highlighted:
  • Li & Liu (2023): LSTM-based forecasts fed into portfolio optimization (max-Sharpe / min-variance frameworks).
  • Liu (2022): constructing financing-constraint indicators and linking them to ESG ratings in the Chinese market.

Implications for AI Economics

• Research agenda:
  • Move from isolated algorithmic gains to system-level evaluation: link forecasting modules, optimizers, and firm-level inference in end-to-end frameworks that reflect economic constraints and institutional realities.
  • Develop loss functions and evaluation protocols that reflect economic objectives (turnover, transaction costs, tail risk, ESG materiality) rather than purely statistical fit.
  • Investigate multi-horizon, multimodal decision architectures that jointly optimize short-term tradability and long-term firm-quality signals (e.g., trading rules that internalize expected ESG-driven financing effects).
• Methodological priorities:
  • Emphasize robust risk estimation and regularization (shrinkage, factor structure) because optimizer sensitivity often dominates realized performance.
  • Favor architectures that combine prediction for ranking/preselection with conservative optimization rules to limit concentration and turnover.
  • Improve interpretability and uncertainty quantification (calibrated predictive intervals, regime-aware models) to support accountability in high-stakes finance and sustainable investment.
• Policy and market design:
  • Recognize ESG measurement heterogeneity; econometric and ML models should model rating noise and cross-provider divergence explicitly before embedding ESG predictions into allocation.
  • Consider governance and accountability for AI-driven allocation—model risk, disclosure standards, and auditability will matter as AI moves from screening to core capital-allocation tools.
• Practical caution:
  • Positive in-sample or sample-specific results for complex models (LSTM, transformers) do not generalize automatically. Economic value requires robustness checks across regimes, transaction-cost inclusion, and alignment between target and decision frequency.
• Opportunity:
  • Integrated AI systems that combine market signals and firm-level sustainability inference can improve investment decisions if designed with explicit constraints, economic targets, and attention to data quality and interpretability.