University of Twente Student Theses

Login

Data-driven retail food waste reduction : a comparison of demand forecasting techniques and dynamic pricing strategies

Felix, P. (2018) Data-driven retail food waste reduction : a comparison of demand forecasting techniques and dynamic pricing strategies.

[img] PDF
3MB
Abstract:Every year one third of all food that is produced is wasted and one of the UN sustainable development goals is to cut food waste in half by 2030. This thesis focuses on two data-driven strategies to reduce perishable food waste at retailers: demand forecasting and dynamic pricing. Improved demand forecasting techniques can prevent excess inventory by better supporting replenishment decisions, whereas dynamic pricing can reduce excess inventory once it exists by stimulating customers to buy older products at a discount. The performance of both traditional and promising new demand forecasting techniques is compared and implementation guidelines are provided for retailers. In addition, simulations were conducted to investigate the performance of different (dynamic) pricing strategies in terms of revenue, waste and stock-outs. For retailers, reducing perishable food waste results in financial and sustainability benefits. Demand Forecasting: Demand forecasting techniques that were included in the performance evaluation are: naive (where the forecast equals sales from last period), exponential smoothing (ES), moving average (MA), linear regression (LINREG), auto-regressive integrated moving average (ARIMA), support vector regression (SVR), multi-layer perceptron (MLP), long-short term memory network (LSTM) and adaptive boost (ADA). These techniques were evaluated based on their performance in forecasting sales for 986 perishable food products from an Ecuadorian supermarket. Performance was measured using the relative root mean squared error measure (RelRMSE), which is a robust measure that indicates how well a certain technique performs relative to the naive forecast. Performance was compared across different forecasting scenarios, which differed in terms of their time detail level, location detail level and horizon. In addition, it was investigated what the influence is of using external factors such as the weather in addition to historical sales data to produce forecasts. Results show that there is no such thing as the ultimate demand forecasting: technique and that performance greatly varies across products and forecasting scenarios. By far the best demand forecasting performance overall can be obtained by automatically selecting the best forecasting technique for each individual product, resulting in (depending on the forecasting scenario) a 6% to 32% RelRMSE improvement compared to always using the naive forecast and a 2% to 10% improvement compared to always using the best individual demand forecasting technique. Adding external factors from the weather, promotion, economic and holiday categories has shown to be valuable for forecasting scenarios that have a daily time detail level, enabling an additional 3% to 9% RelRMSE improvement compared to using historical sales data only. For each forecasting scenario, detailed results and an overview of the top 3 best performing individual techniques are provided in this thesis to help retailers select the appropriate demand forecasting technique(s) in their situation. In addition, a process is provided with step-by-step guidelines for retailers on how to improve the wider demand forecasting process. It not only considers how to select the right demand forecasting technique(s), which form the quantitative core of the demand forecasting process, but also discusses how to assess forecasting capabilities and which qualitative factors should be taken into account, such as implementation in decision support systems, adoption factors and organizational factors. Dynamic Pricing: A simulation study was conducted to determine which (dynamic) pricing strategy for perishable products performs best in terms of total revenue, waste and stock-outs. The simulation considers a monopolist grocer selling a single perishable product with a fixed shelf life. The product can be replenished each day based on the demand forecast and a safety factor. Customers have different characteristics: some customers are regular customers that only pay attention to price, whereas others are date-checking customers that also pay attention to remaining shelf life and aim to choose a product that maximizes their value-for-money. Four main pricing strategies for marking down perishable products that approach their expiry date were compared. Strategy 1 applies no price changes at all and serves as a baseline. Strategy 2 applies a fixed discount D at the last day before expiry, while strategy 3 spreads that discount over the last S days before expiry. Strategy 4 dynamically determines discount percentages based on the demand forecast and the remaining inventory. Almost all strategies that applied a discount resulted in a waste reduction, but they regularly resulted in significantly lower revenues. Multiple experiments were conducted to investigate the effects of varying assumptions for simulation settings. Discounting was most beneficial when product demand was more elastic and when more customers checked expiry dates, because that resulted in higher waste reductions and less negative (or even positive) changes to revenue at the same time. The fixed pricing strategy that most frequently performed best was strategy 2 with a fixed discount of 20% on the last day before the expiration date. Surprisingly, the dynamic pricing strategies did not always perform better than a fixed price strategy, which could be due to the fact that these strategies relied upon imperfect demand forecasts and hence their estimated optimal discount percentage might have been off base. A dynamic pricing strategy already outperformed the best fixed strategy when initial waste levels for a product were high or when a large percentage of customers were regular customers. Conclusion and Discussion: Both demand forecasting process improvement and dynamic pricing strategies for marking down products that near their expiry dates have a positive impact on waste reduction. Grocers are advised to initially focus on improving the demand forecasting process, since that reduces waste, but does not have the negative impact on revenue that discounting strategies frequently have. Preventing inventory excesses from occurring through improved demand forecasting is better than trying to resolve such excesses by discounting products that approach their expiration dates. Grocers are advised to follow the demand forecasting improvement process, to use this study as a benchmark for forecasting technique selection, to implement multiple forecasting techniques and to automatically select the best forecasting technique for each product. To resolve any excesses that do occur in stores, grocers are advised to use the pricing strategy that showed best performance in similar situations in the simulations. The best fixed strategy in general is applying a 20% fixed discount on the last day before expiry. A dynamic pricing strategy only outperformed a fixed strategy in a few situations. This study makes several contributions to both retail practice and the scientific fields of demand forecasting and dynamic pricing. First of all, this study provides a robust and objective comparison of forecasting techniques in different scenarios within a food retail context. To the best of our knowledge, it is the first study that conducts such a comparison. Results show the impact of automatically selecting the best forecasting techniques for each individual product and for including external factors and can be used by retailers as a benchmark for forecasting technique selection. In addition, this study proposes a process for demand forecasting improvement and forecasting technique selection, which can guide retailers in their demand forecasting improvement efforts. This study also provides a robust performance comparison of different pricing strategies that mark down perishable products, giving retailers insight into the impact on revenue, waste and stock-outs. In addition, tools were developed in the form of an algorithm for automatic best DF technique selection (and configuration) and a pricing simulation that can be reused in future work.
Item Type:Essay (Master)
Clients:
University of Twente, Enschede, The Netherlands
Deloitte, Amsterdam, The Netherlands
Faculty:EEMCS: Electrical Engineering, Mathematics and Computer Science
Subject:54 computer science, 85 business administration, organizational science
Programme:Computer Science MSc (60300)
Link to this item:https://purl.utwente.nl/essays/76370
Export this item as:BibTeX
EndNote
HTML Citation
Reference Manager

 

Repository Staff Only: item control page