Anomaly Detection and Day-of-the-Week Forecasting of NSE NIFTY Using a Hybridized Neural Network

Abstract

Financial markets, particularly stock markets, exhibit complex behaviors. Calendar anomalies and sudden structural changes pose challenges to standard forecasting models. Linear models cannot predict these changes effectively. One significant calendar anomaly is the day-of-the-week (DoW) effect, which has been particularly studied in emerging markets such as India. This study provides empirical evidence that explores the DoW anomalies in the NSE NIFTY
50 index and proposes a hybrid neural network framework that combines anomaly detection with daily forecasting. This study employs unsupervised anomaly detection using an autoencoder neural network to identify unusual trading days. These indicators were then included in a CNN-LSTM-Attention model for return forecasting. The results provide evidence of significant day-of-the-week effects on the NSE NIFTY 50 index. The proposed hybrid autoencoder-CNN-LSTM-attention framework achieved superior forecasting performance compared with conventional ARIMA and standalone LSTM models, reducing the RMSE by approximately 31.7% (from 1.42 to 0.97) and improving the directional accuracy from 52% to 63%. Future research should incorporate sentiment indicators, macroeconomic variables, and cross-market comparisons to further enhance forecasting performance.