Causal Observables for Financial Forecasting

This paper addresses the challenge of short-horizon forecasting in financial markets, focusing on the construction of interpretable and causal signals. It moves beyond direct price prediction and instead concentrates on building a composite observable from micro-features, emphasizing online computability and causal constraints. The methodology involves causal centering, linear aggregation, Kalman filtering, and an adaptive forward-like operator. The study's significance lies in its focus on interpretability and causal design within the context of non-stationary markets, a crucial aspect for real-world financial applications. The paper's limitations are also highlighted, acknowledging the challenges of regime shifts.