Incorporating causal notions to forecasting time series: a case study
The paper proposes a causal inference framework integrating treatment effects with machine learning to enhance financial time series forecasting.
What it examines
This study proposes a new forecasting framework where causal reasoning, via treatment (positive returns) and control (negative returns), is integrated into financial time series forecasting. It combines econometric and machine learning models to capture directional changes and aims to improve prediction accuracy over standard methods.
What it concludes
The proposed causal forecasting framework improves traditional models by selecting machine learning models based on return direction. It shows robust, statistically significant performance in financial forecasting. Applications include stock prediction, portfolio management, and related fields. Future work may extend its causal approach to other complex data settings.
Evidence objects
A novel causal forecasting framework trains distinct models for each return direction, applying rolling windows, hyperparameter optimization, and uplift modeling; a decision module selects optimal forecasts based on predicted directionality.
key_findings bullet 2 · key_findings · validation V0
Results reveal minor $MSE$ improvements as statistically significant via Model Confidence Set tests, with performance offset by quarterly fluctuations and modest gains, underscoring the need for research on complex categorizations.
key_findings bullet 3 · key_findings · validation V0
Researchers merge causal inference and financial forecasting by dividing data into treatment (positive returns) and control (negative returns), enabling LSTMs to yield improved predictions over ARIMA and Random Walk models.
key_findings bullet 1 · key_findings · validation V0
Integrating causal treatment/control concepts with \$return\$ sign analysis, this paper presents a novel framework for financial time series forecasting. It demonstrates potential improvements in market prediction and portfolio optimization over conventional econometric and machine learning models. Though echoing recent work, its moderately original approach offers fresh and compelling analytical insights.
key_findings bullet 4 · key_findings · validation V0
Raw abstract and provenance
- … time series have been analyzed with a wide variety of models and approaches, some of which can forecast with … into a forecasting framework to better predict financial time …
Source row: 1083 · abstract type: snippet