k-Nearest Neighbours, a Novel Machine Learning Method for Retail Promotional Forecasting

Rodriguez-Calderon, Carlos-Eduardo and Crone, Sven and Kourentzes, Nikos and Sachs, Anna-Lena and Svetunkov, Ivan and Barrow, Devon and Fildes, Robert (2026) k-Nearest Neighbours, a Novel Machine Learning Method for Retail Promotional Forecasting. PhD thesis, Lancaster University.

Text (2026RodriguezCalderonPhD)
2026RodriguezCalderonPhD.pdf - Published Version
Available under License Creative Commons Attribution-NonCommercial-NoDerivs.
Download (6MB)

Abstract

Forecasting accuracy remains a persistent challenge in competitive markets with widespread use of promotions. Machine learning algorithms have emerged as an alternative to traditional (statistical) ones to improve forecast accuracy; however, they frequently present an obscure prediction logic when attempting to trace the origin of the results. This thesis investigates whether averages, a foundational approach approved by forecasters, can underpin classical forecasting when implemented through k-NN, an explainable machine learning algorithm. To understand practitioner needs, a survey captures current forecaster preferences, building on insights from an initial study conducted a decade earlier. Findings illustrate continued reliance on averages and inform the empirical evaluation of k-NN for promotional forecasting, focusing on feature preprocessing. The methodology combines survey results with a quantitative analysis of k-NN performance on data from an undisclosed retailer/wholesaler. The literature review highlights a gap in independent studies that represent demand planners’ perspectives, as well as a scarcity of research on the application of k-NN regression for retail forecasting. This motivates the primary research question: Can averaging-based machine learning algorithms like k-NN deliver competitive forecasts amid statistical algorithms, and how do transformation and autoregressive features affect predictions? Key contributions include: (1) updated insights into forecasters’ methodological preferences, (2) empirical validation of k-NN as an alternative for demand forecasting, and (3) practical guidance on transforming feature vectors and lagged values to enhance forecast accuracy. These results bridge theory and practice, providing actionable recommendations for researchers and practitioners. Future work may extend the evaluation to further transformations or feature discrimination and explore hybrid forecast approaches to capture extra components.