On Issuance: What Matters, What Doesn’t and Where Green Fits

Abstract

Green bond pricing and the greenium remains unsettled, and most evidence comes from secondary markets where liquidity and trading frictions complicate interpretation. This thesis focuses on the primary market and asks two practical questions: whether the green label is associated with an issuance G-spread differential once standard spread determinants are controlled for, and whether machine-learning models improve out-of-sample prediction. The dataset consists of corporate bonds issued between 2023 and 2025 and totals 3654 bonds. Issuance G-spread is used as the dependent variable and controls include bond characteristics and market conditions, such as broad credit spreads and volatility proxies. Issuance G-spreads are estimated with a nested OLS ladder to track how the green coefficient behaves as control blocks are introduced. Predictive performance is benchmarked across OLS, LASSO, and Random Forest and evaluated on a held-out test set, across 100 repeated random splits and in an out-of-time setting that trains on earlier years and tests on a later period. The results show that green labelled issues are associated with a spread difference of about +16.4 bps at the representative spread level. Random Forest provides a large improvement in predictive accuracy relative to linear benchmarks. On the held-out test set, RMSE decreases from 0.251 to 0.152 and test R² increases from 0.858 to 0.948. In out-of-time evaluation, performance falls for all models but Random Forest remains most accurate. State splits indicate that implied green differential is larger in tight conditions at 28–31 bps and smaller and less precise in stressed regimes at 9–10 bps. Issuance spreads are dominated by conventional drivers such as credit quality, deal structure, and issuance-time market conditions. The green label adds secondary but sometimes economically meaningful information, and the magnitude is regime dependent.