Decoding financial markets: Empirical DGPs as the key to model selection and forecasting excellence – A proof of concept

In this study we demonstrate whether information about the empirical data generating process (DGP) can optimise quantitative model selection and present a proof-of-concept. We derive the empirical DGP characteristics of nine (financial) time-series and interpret these as model requirements. These insights are tested via a comparative cascadic out-of-sample prediction scheme to demonstrate potential outperformance. The empirical DGP characteristics of (denoised) daily adjusted logarithmic returns are extracted with a nonlinear dynamics analysis framework. Thereinafter, various forecasting models and error metrics are implemented for the non-filtered returns. The models’ out-of-sample performance is subsequently ranked across these multiple metrics. Finally, we assess the models' alignment with the empirical DGPs. Our results show that all time-series exhibit very similar dynamics, independent of the underlying asset class. The dynamics are characterised by a mixture of deterministic (hyper-)chaotic and stochastically imbued quasi-periodic motions. Further, these dynamics are visualised through the reconstructions of the strange (fractal) attractors of the time-series. Models that capture the dynamics best outperform standard benchmarks and other state-of-the-art methods. Nonetheless, we have to state an overall lack of suitable DGP-conform models. Finally, we critically discuss the robustness, sensitivity and implications of our approach and findings.