A Transformer-Based Explainable Machine Learning Framework for Heterogeneous ESG Risk and Return Prediction Using Text Intelligence (Future Technology)

Summary

• HET-ESGFormer is evaluated on 3,182 U.S.-listed firms over 2015 to 2024 using 10-K filings, sustainability reports, earnings call transcripts and ESG-tagged news.p.1
• Against nine baselines, it reduces return RMSE from 0.0417 to 0.0389 and QLIKE from 0.148 to 0.128.p.1
• Zeroing the text channel shows that it reduces QLIKE from 0.169 to 0.128, a relative improvement of 24.3%.p.8
• The monthly decile long-short portfolio with a 10-bp one-way cost earns 11.4% annually, a Sharpe ratio of 1.31, and an 18.7% maximum drawdown.p.10
• The VaR backtest at 95% coverage yields a violation frequency of 5.3%.p.10
• The sample contains 28,647 firm-year observations and 1,842,617 ESG news items.p.6

Key insights

1. At a rationale size of k = 10%, the dual-channel explanation achieves comprehensiveness of 0.297 and sufficiency of 0.089.p.10
2. At α = 0.05 the pinball loss drops from 0.0219 to 0.0186, an improvement of 15.1%.p.8
3. The text channel raises R2OOS from 0.055 to 0.083, an increase of 2.8 percentage points.p.8
4. In the COVID-19 and 2022 rate-hike subsamples, HET-ESGFormer still outperforms Longformer by 21.9% and 16.8%, respectively.p.12
5. QLIKE remains between 0.121 and 0.146 as the horizon varies from 1 to 12 months.p.12
6. Vendor-rating portfolios have Sharpe ratios of 0.74 and 0.62.p.10