Abstract—Stock Price Prediction is a complex challenge in
the field of finance due to its inherent volatility and its
non-linear, and non-stationary nature. Current Statistical and
Machine Learning models struggle to capture non-linear and
non stationary patterns. Stacked ensembles improve prediction
performance but struggle with distribution shifts and poor
extrapolation when the market moves beyond the training range.
In this study, we train and evaluate three base models (LSTM,
CNN, and TCN) on historical NIFTY 50 OHLC data sourced
from Yahoo Finance. These are combined into two ensemble
architectures: LSTM+CNN and LSTM+TCN. Their predictions
are transformed into delta space, and then 13 meta-learners

are systematically evaluated to find the most effective ensem-
ble architecture.Furthermore, three feature scaling techniques:

MinMax, Standard, and Robust Scaling are compared, and the
robustness of the proposed framework is evaluated using five-fold
cross-validation. Experimental results show improved robustness,
higher R
2
and lower prediction errors.

Terms—Stock Price Forecasting, Hybrid Ensemble Learning , Delta-Transformation,NIFTY-50, LSTM, CNN, TCN, Time Series Forecasting