Data Aggregation and Demand Prediction

Abstract

High accuracy in demand prediction allows retailers to effectively manage their inventory and mitigate stock-outs and excess supply. A typical retail setting involves predicting the demand for hundreds of items simultaneously, some with abundant historical data and others with scarce data. In “Data Aggregation and Demand Prediction,” Cohen, Zhang, and Jiao propose a novel practical method, called data aggregation with clustering (DAC), which balances the tradeoff between data aggregation and model flexibility. DAC empowers retailers to predict demand while optimally identifying the features that should be estimated at the item, cluster, and aggregate levels. Theoretically, DAC yields a consistent estimate, along with improved prediction errors relative to the benchmark that estimates a different model for each item. Practically, DAC yields a higher demand prediction accuracy relative to many common benchmarks using a real data set from a large online retailer.

New product life cycle curve modeling and forecasting with product attributes and promotion: A Bayesian functional approach

“…Other valuable covariate information, such as leading-indicator products, distribution channel, advance purchases, social media, and Google trend, has also been exploited in sales forecasting (Aytac & Wu, 2013;Boone et al, 2018;Cui et al, 2018;Moe & Fader, 2002;Neelamegham & Chintagunta, 1999). In addition, clustering techniques and machine learning models have been leveraged to boost forecasting performance of new product sales (Baardman et al, 2017;Cohen et al, 2022b;Ferreira et al, 2016); the readers may find a comprehensive and recent review of forecasting techniques in Cohen et al (2022a). Unlike the majority of the sales forecasting literature that focuses on scalar target variables, our work presents a functional regression framework to predict sales curves as a whole, which captures the temporal sales pattern within a PLC explicitly.…”

Automated update tools to augment the wisdom of crowds in geopolitical forecasting.

“…Machine models outperform humans (including experts) in many tasks involving predictive or diagnostic judgments (Cowgill, 2017; Dawes, 1979; Grove et al, 2000; Meehl, 1954), particularly in stable environments (Kremer et al, 2011). However, developing formal forecasting models is challenging (e.g., because of having many features but little data, Cohen et al, 2022) and most existing approaches rely on heuristics, past experience, or trial and error (Lawrence et al, 2006). Furthermore, forecasts are not static but rather benefit from frequent updates (Atanasov et al, 2020; Mellers, Stone, Atanasov, et al, 2015; Mellers, Stone, Murray, et al, 2015), increasing the costs of using human crowds for forecasting in unstable contexts over longer time periods.…”