Back to search

CSUniversity San Bernardino

EVALUATING MACHINE LEARNING DAILY SALES FORECASTING FOR SMALL LOCAL FARMS: A CASE STUDY OF ORANGE SALES AT ARNOTT FARMS

Abstract

dc:description.abstract

Of the many methods that can be used to stabilize a supply chain, obtaining an accurate demand forecast is described in the literature as being a key step. With advances in computing power, machine learning-based demand forecasting models have emerged as an alternative to traditional statistics-based methods. Arnott Farms is a small farm in Mentone, California that operates a small grocery store as a distribution channel. For this research, I collected point-of-sale data from the store, restructured and augmented it with engineered feature variables and outside data, and developed various machine learning algorithms to evaluate whether a small farm can utilize machine learning to forecast product sales and aid in decision-making for allocating harvested goods between the store and other distribution channels. The research questions are: RQ1: What machine learning models are most accurate for predicting sales when used on historical sales data in the context of a small farm?; RQ2: Which combinations of input variables, when added or removed to historical sales data, have the largest net impact on model accuracy?; RQ3: What external data, such as weather data and gas price data, when added or removed to historical sales data, have the largest net impact on model accuracy?; and RQ4: How do specific machine learning models compare in accuracy to traditional statistics-based time series forecasting methods? I used Python and the scikit-learn library to train and test predictive models to forecast sales of milk and 25 lb. bags of oranges. I compared models using the same dataset and evaluation metrics. I evaluated the impact of different input variables on the accuracy of the models by running them with and without variables to test the change in results. I found that, for this farm, some forms of outside data, particularly precipitation, did not improve forecast accuracy. Rolling averages were effective as an engineered model feature. The principal finding of the study was that, for this business context, a simple seven-day moving average yielded comparable forecast accuracy to advanced machine learning models. In practice, this finding supports small farmers who are searching for a way to implement data-driven forecasting into their business workflow and can benefit from research specific to their business context. An area for future study can be to assess the value of social media data from a business-owned social media account as an outside data input into sales forecasting models.

Degree

thesis:*
Name thesis:degree_name
Master of Science in Information Systems and Technology
Level thesis:degree_level
Thesis
Discipline thesis:degree_discipline
Information and Decision Sciences

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Strong, Ryan Lawrence
Contributors dc:contributor
  • Molavi, Nima

Subjects

dc:subject × 13

Identifiers

dc:identifier.*
Repository record dc:identifier
https://scholarworks.lib.csusb.edu/etd/2553

Chain of custody

source
Harvested from
CSUniversity San Bernardino
Base URL
scholarworks.lib.csusb.edu/do/oai/
Last updated
2026-11-20
Source record
OAI-PMH GetRecord
citation

Strong, Ryan Lawrence. EVALUATING MACHINE LEARNING DAILY SALES FORECASTING FOR SMALL LOCAL FARMS: A CASE STUDY OF ORANGE SALES AT ARNOTT FARMS. Thesis thesis, https://scholarworks.lib.csusb.edu/etd/2553