Dynamic Multi-Criteria Portfolio Selection Integrating Transformer-Based Financial Forecasting with Peer-Prediction Trees

Portfolio optimization demands simultaneous consideration of multiple conflicting criteria under uncertainty, yet prevailing approaches either rely on some black-box machine learning (ML) models that sacrifice interpretability or employ classical multi-criteria decision-making (MCDM) methods lacking predictive foresight. This paper proposes a two-stage framework integrating a Transformer encoder for multi-output financial forecasting with the Peer-Prediction Trees for MCDM (PPT-MCDM) method for dynamic asset ranking and portfolio construction. The Transformer generates forward-looking predictions of next-period return, volatility, and maximum drawdown, while PPT-MCDM ranks assets by their excess performance index (EPI), measuring how much each asset’s multi-criteria profile exceeds data-driven peer expectations. The framework is validated on 28 sector and thematic exchange-traded funds (ETFs) over a 51-month out-of-sample period from January 2022 to March 2026. The PPT-MCDM portfolio achieves an annualized return of 11.99% with a Sharpe ratio of 0.589 and maximum drawdown of 18.80%, compared to the S&P 500 benchmark delivering 9.16% return, Sharpe ratio of 0.391, and maximum drawdown of 20.25%. An ablation study confirms that Transformer predictions improve the Sharpe ratio by 39.9% relative to using only observed backward-looking criteria. The main contributions of this work are three-fold: first, the development of a two-stage framework integrating deep learning forecasting with interpretable MCDM-based portfolio ranking; second, the first application of PPT-MCDM method to dynamic portfolio optimization with expanding-window retraining; third, empirical evidence that the framework outperforms the S&P 500 on both return and risk-adjusted metrics during a period encompassing both bear and bull market conditions.