A BiLSTM–NNAR–Tree Based Stacked Ensemble for Financial Forecasting

Abstract

Forecasting stock prices is hard. Prices are nonlinear. They depend on past values. Market behavior also changes over time. Sudden events can shift the trend. Because of this, old statistical models often miss key patterns. Single machine learning or deep learning models also have their own weak points. To handle this, we propose a stacked ensemble model to forecast the next closing price directly. We work with the original price series. We do
not difference the data. This keeps the values in the real money scale. We use three different models. A BiLSTM learns long-range patterns from both past and future context in the sequence. An NNAR model captures short-term nonlinear effects using past price lags. Then a tree-based gradient boosting model learns from the remaining errors. It helps
correct mistakes by catching rule-like changes, such as regime shifts and threshold effects. Finally, we combine the outputs of these models using stacking. A simple linear model acts as the final meta-model. This makes the final forecast more stable and easier to explain. We test the method on daily closing prices of major pharma and automobile stocks in India from 2017 to 2025. This period includes different market phases, including the COVID-19 shock. The results show that our ensemble gives better forecasts than common statistical models, machine learning models, and deep learning models. It performs better on RMSE, MAE, MAPE, and R2, which suggests it is reliable for real market data.