Real Time Sales Forecasting in Omnichannel Retail Using a Hadoop Based Hybrid CNN–LSTM Deep Learning Framework

Abstract

In an omnichannel retail environment, accurate real time sales forecasts are critical for inventory optimisation and dynamic pricing. This study proposes a Hadoop based hybrid CNN–LSTM deep learning framework that leverages Hadoop’s distributed computing capabilities to process almost 20 million multi source sales records collected over two years. The convolutional layers and recurrent layers cooperate to capture local pulses and long range dependencies, respectively. Systematic experiments show that, compared with classical ARIMA and various machine learning baselines, the proposed model reduces the mean squared error (MSE) by approximately 45 % and increases the coefficient of determination (R²) by about 15 %. Within randomly selected 30 day windows, the model stably tracks high frequency intra week fluctuations while effectively suppressing noise spikes. Moreover, the Hadoop cluster shortens the total training time from 14 h to 3.5 h and compresses single inference latency to 48 ms, satisfying second level business decision requirements. Ablation studies further verify the complementary benefits of the convolutional and recurrent components; removing either leads to significant performance degradation. After deployment at a partner retailer, the stock out rate and dead inventory were reduced by 7.8 % and 6.1 %, respectively, demonstrating the commercial value of the approach. Limitations include cold start bias for new items, underestimation of extreme promotion peaks and insufficient model interpretability. Future work will explore graph convolution to incorporate spatial correlations, self supervised pre training to alleviate cold starts and attention mechanisms to enhance interpretability—thus driving retail sales forecasting towards greater accuracy, trustworthiness and inclusiveness.