Deep-learning models, led by GRUs, forecast a simulated Thai ESG stock index substantially better than classical ARIMA-style approaches over a 36-day horizon. The advantage endures even with limited historical data, implying advanced AI can strengthen ESG market signals in data-constrained emerging markets.

Main Finding

Deep learning models substantially outperform traditional statistical time-series methods in forecasting an ESG-based stock index constructed for Thailand; among the tested architectures, GRU achieved the highest accuracy and the most consistent robustness when training data were limited. This demonstrates that advanced AI techniques better capture nonlinear ESG market dynamics in data-constrained emerging-market settings.

Key Points

• Context: Forecasting ESG-based indices is challenging in emerging markets due to limited historical data and index construction specifics.
• Target: A realistic simulated SET ESG Index was constructed using free‑float–adjusted market capitalization and semiannual rebalancing rules mirroring the Stock Exchange of Thailand.
• Forecast horizon: Multi-step out-of-sample prediction over 36 trading days.
• Models compared:
  • Traditional statistical: ARIMA, SARIMA, SARIMAX.
  • Deep learning architectures: vanilla RNN, GRU, LSTM, DF-RNN, DeepAR, DSSM, Deep Renewal.
• Performance: All deep learning models outperformed statistical models on multi-step forecasts; GRU delivered the best accuracy and most consistent robustness across reduced-data scenarios.
• Robustness: Deep models—particularly GRU—maintained performance when historical data were scarce, highlighting suitability for emerging/evolving markets.
• Reproducibility: The study provides a replicable modeling pipeline (index construction + training/evaluation setup) for ESG index forecasting under data constraints.

Data & Methods

• Index construction: Simulated SET ESG Index built from Thai equity universe using free‑float–adjusted market-cap weights and semiannual rebalancing consistent with Stock Exchange of Thailand methodology.
• Data characteristics: Limited historical coverage to emulate typical emerging-market ESG data availability.
• Forecasting setup:
  • Multi-step prediction task with a 36-day forecast horizon.
  • Models trained and validated on the constructed index time series; reduced-data experiments simulated varying data availability to test robustness.
• Models and training:
  • Classical time-series baselines: ARIMA, SARIMA, SARIMAX (for exogenous inputs).
  • Deep learning baselines: recurrent and probabilistic forecasting models (RNN, GRU, LSTM), plus DF‑RNN, DeepAR, DSSM, Deep Renewal.
  • Evaluation used standard forecasting metrics (e.g., RMSE, MAPE and related multi-step error measures) and robustness checks across reduced-data regimes.
• Key methodological emphasis: fair comparison across models with identical train/validation/test splits and rebalancing-consistent index construction to reflect real-world trading rules.

Implications for AI Economics

• For sustainable finance and practitioners:
  • AI (especially GRU-like recurrent architectures) can improve short-term ESG index forecasting in emerging markets, aiding portfolio construction, timing, and tactical allocation for ESG-focused funds.
  • Improved forecasts support risk management (short-term volatility and drawdown anticipation) and liquidity planning around semiannual rebalancings.
• For market participants and regulators:
  • Better forecasting tools can increase market efficiency for ESG-labelled products in markets with sparse historical data, potentially driving greater investor confidence and capital flows into sustainable investment strategies.
  • Regulators and exchanges should note the value of transparent index methodology and timely ESG disclosures to improve model inputs and market stability.
• For researchers and modelers:
  • Demonstrates that deep learning can capture nonlinearities in ESG-driven price dynamics that linear time-series models miss; GRU is a strong default choice under data constraints.
  • The provided, reproducible pipeline enables transfer and adaptation to other emerging markets; promising extensions include incorporation of alternative data (news, sentiment, supply-chain ESG signals), transfer learning across markets, and model explainability to meet regulatory and investor transparency needs.
• Caveats & practical considerations:
  • Deep models require careful cross-validation, regularization, and computational resources; overfitting and data-snooping risks remain, especially with limited data.
  • Model interpretability and auditability are important for institutional adoption—techniques for explainability and stress-testing should accompany deployment.
• Policy-relevant note: Enhancing ESG data coverage and standardization in emerging markets will amplify the value of AI forecasting and support resilient sustainable finance ecosystems.