Analise preditiva dos spreads de crédito no mercado de debêntures do IDEX-CDI JGP por meio de machine learning

Abstract

The Brazilian private credit market has gained increasing relevance as an alternative source of corporate financing, driven by the rise in debenture issuances and the growing role of fixed income investment funds. In this context, understanding the dynamics of credit spreads and their relationship with net fund flows becomes essential for portfolio management. This study analyzes this relationship using the Idex–CDI JGP, a representative index of liquid debentures indexed to the CDI, as a benchmark for measuring spread behavior. The methodology combines econometric and machine learning techniques, beginning with a correlation analysis between spreads and net fund flows using different lags and deseasonalized series. A moderate negative short-term correlation was observed, consistent with higher risk aversion, but it strengthens significantly after deseasonalization, reaching a Pearson coefficient above 0.82 at around a ten-month lag. The Weighted Least Squares (WLS) model delivered an excellent fit, with an R² above 0.90, and showed that the lagged spread is a strong determinant of net fund flows. To forecast the spreads, both univariate models (spread only) and multivariate models (spread and net fund flows) were tested within a three-stage pipeline: an autoregressive model with Elastic Net, a correction step, and a final horizon-based forecasting stage using the HistGradientBoostingRegressor, along with Spearman selection, clipping, and calibration. The univariate model performed strongly in the very short term but lost accuracy over longer horizons. The incorporation of flow-related variables produced limited gains at short horizons, but it improved predictive accuracy over longer forecasting windows, indicating that net fund flows contribute to identifying different phases of the credit cycle.