Advanced Deep Learning Techniques for USD/MXN Structured Notes Predictive Modeling

Abstract

This thesis investigates whether modern neural networks can enhance the issuance and selection of capital-protected structured notes in the Mexican market, concentrating on instruments linked to the USD/MXN exchange rate. Standard feed-forward and shallow recurrent models were first applied to a broad spectrum of payoff profiles—range accruals, double no-touch, and directional notes—but delivered limited out-of-sample stability and negligible economic value. Reframing the problem as a binary 7-day USD/MXN direction forecast narrowed the note universe to Strike Up and Strike Down structures and allowed the training set to extend back to 2010. A dual-stream GRU architecture was developed in which technical indicators (momentum, oscillators) and fundamental drivers (rate differentials, macro surprises) flow through parallel recurrent branches before merging in a shared decision layer. Bayesian hyper-parameter optimisation selected a compact configuration (five layers, 372 hidden units, dropout 0.3). When evaluated on 2024 out-of-sample data, the model achieved Accuracy = 64.7%, F1 = 71.7%, and AUC = 0.60, comfortably outperforming ARIMAX, logistic regression, and decision-tree baselines. The cumulative true-positive excess return (TPefe) tracked the CETE-28 benchmark within approximately ten basis points after embedded structuring fees, demonstrating commercial viability for investors who require weekly liquidity, principal protection, and yields at least equal to the domestic risk-free rate. The study shows that separating technical and fundamental information into dedicated recurrent channels, coupled with gated-unit dynamics and commission-aware evaluation, provides a robust and scalable template for machine-learning-driven structured-note selection in Mexico’s evolving markets.