Authors: Vansh Shisodia, Saibee Alam, Anish Kushwaha, Aarchi Goyal
Abstract: This research constructs a hybrid system for one- step-ahead (H=1) stock forecasting, addressing the non-linear and non-stationary nature of financial time-series. The objective is twofold: a regression task for the Adjusted Close Price and a binary classification task for directional movement. The pro- posed ensemble design combines three model families: classical econometrics (ARIMA), deep learning (LSTM) for long-term dependencies, and ensemble tree methods (XGBoost, RF) for non- linear feature interactions. The methodology emphasizes rigorous feature engineering, including technical indicators and GARCH- derived volatility features, and robust validation using Time Series Cross-Validation (TSCV) and Nested Cross-Validation (nCV). The system culminates in a stacked ensemble (blending layer) and utilizes advanced loss functions like Huber Loss to manage heavy-tailed return distributions Evaluation is based on both statistical fit and financial utility metrics, such as directional accuracy and the Sharpe Ratio.
DOI:
