Georgios Theodoridis

Accurate retail demand forecasting is integral to the operational efficiency of any retail business. As demand is described over time, the prediction of demand is a time-series forecasting problem which may be addressed in a univariate manner, via statistical methods and simplistic machine learning approaches, or in a multivariate fashion using generic deep learning forecasters that are well-established in other fields. This study analyzes, optimizes, trains and tests such forecasters, namely the Temporal Fusion Transformer and the Temporal Convolutional Network, alongside the recently proposed Time-Series Mixer, to accurately forecast retail demand given a dataset of historical sales in 45 stores with their accompanied features. Moreover, the present work proposes a novel extension of the Time-Series Mixer architecture, the LSTMixer, which utilizes an additional Long Short-Term Memory block to achieve better forecasts. The results indicate that the proposed LSTMixer model is the better predictor, whilst all the other aforementioned models outperform the common statistical and machine learning methods. An ablation test is also performed to ensure that the extension within the LSTMixer design is responsible for the improved results. The findings promote the use of deep learning models for retail demand forecasting problems and establish LSTMixer as a viable and efficient option.

Retail demand forecasting is a highly intricate and multilevel problem as it may include numerous products, under multiple categories, and potentially sold in many different stores. Forecasts may also process products in isolation or collectively so as to detect mutual correlations. In technical terms, retail demand forecasting may be broken down to a multivariate time series forecasting problem where the time axis is prevalent and introduces added complexity which necessitates proper usage of advanced methodologies with machine/deep learning backgrounds. Popular boosting regressors, such as XGBoost and LightGBM, are excellent machine learning candidates with extensive bibliographic backgrounds covering multiple scientific fields and applications including generic multivariate time series forecasts. The Temporal Convolutional Network and the Temporal Fusion Transformer are recently proposed deep neural network architectures that found success in specific multivariate scenarios but are yet to be tested in a variety of related fields such as retail demand forecasting. Therefore, within this paper, these novel in-the-field models are compared to the aforementioned boosting approaches alongside popular statistical univariate methods, namely Exponential Smoothing and SARIMA. To properly compare the selected models and attempt to generalize their usability, two separate datasets are analyzed and forecasted; an item sales dataset with 500 time series and a category sales dataset with 540 time series. The findings suggest that the deep learning solutions are the better predictors, with the Transformer model surpassing the Boosting solutions by up to 16.8% sMAPE decrease and the statistical approaches by up to 28.6% decrease, further substantiating the notion that deep learning techniques are exceptionally promising in handling large-scale, non-linear and outlier data whilst suggesting that retail demand forecasting does benefit from multivariate approaches and the application of advanced machine learning methods.

A low complexity MIMO Volterra Decision Feedback Equalizers (VDFE) for optical transmission systems employing NRZ-DQPSK signalling is proposed. Based on a comparative study performed by means of simulations, it is proved that the proposed equalizers maintain the required efficiency in terms of BER, achieving significant reduction in terms of complexity. Also, suitable architectures for high-speed FPGA implementations are presented. A 8-input 2-output low complexity VDFE involving three taps feed-forward filtering and two taps backward filtering was implemented on a single state-of-the-art FPGA. The target rate of 40 Gb/s is achieved by applying extensive pipelining and parallelism and fully exploiting specific FPGA features.