Hybrid DB3 Wavelet Decomposition with Levenberg-Marquardt Neural Network for Stock Price Forecasting: Evidence from ICICI Bank, India

Abstract

Accurate prediction of equity price movements remains a persistent challenge due to the non-stationary, noisy, and chaotic nature of financial time series. This study presents a hybrid forecasting framework combining Discrete Wavelet Transform (DWT) using the DB3 mother wavelet with a Levenberg-Marquardt (LM) algorithm optimized backpropagation neural network. The DB3 wavelet decomposes original price signals into three detail levels and one approximation component, isolating structured patterns from stochastic noise. A feedforward neural network with architecture [25,15] was trained using the LM backpropagation algorithm. Evaluation on daily trading data of ICICI Bank (NSE) over a one-year out-of-sample period (27 May 2025 to 26 May 2026) yielded RMSE of 1.8534%, MSE of 0.0345%, Efficiency of 98.15%, and MAPE of 1.80%. The 7-day forecast from 27 May 2026 indicated a -4.63% decline. The DB3-LM hybrid achieved perfect inverse correlation (-1.000) between RMSE and Efficiency across architectures, and near-unity positive correlation (0.996) between MSE and RMSE. This research demonstrates that DB3 wavelet denoising coupled with LM-optimized compact networks outperforms deeper architectures for Indian stock market forecasting.