Skip to main content

LUP Student Papers

LUND UNIVERSITY LIBRARIES

Machine Learning as a Tool for Demand Forecasting: A Study at HMS Networks

Jhaveri, Jacob Filip LU and Skogman, Måns LU (2026) MTTM05 20261
Production Management
Engineering Logistics
Abstract
This thesis investigates the applicability of using machine learning for demand
forecasting at HMS Networks. Using the design science research methodology, multiple forecasting artifacts were developed and tested iteratively, using
historical sales data. Four tree-based models; Decision Tree, Random Forest,
LightGBM and XGBoost, were tested with different pre-processing and feature
engineering techniques such as first-order differencing, scaling and varying levels of granularity. The results show that performance was primarily improved
through better data representations, rather than through model selection alone.
Particularly, first-order difference and scaling improved the models’ abilities to
generalize and reduced forecasting... (More)
This thesis investigates the applicability of using machine learning for demand
forecasting at HMS Networks. Using the design science research methodology, multiple forecasting artifacts were developed and tested iteratively, using
historical sales data. Four tree-based models; Decision Tree, Random Forest,
LightGBM and XGBoost, were tested with different pre-processing and feature
engineering techniques such as first-order differencing, scaling and varying levels of granularity. The results show that performance was primarily improved
through better data representations, rather than through model selection alone.
Particularly, first-order difference and scaling improved the models’ abilities to
generalize and reduced forecasting bias. Furthermore, the temporal granularity
was shown to strongly affect the stability of the forecast, where monthly granularity generated more stable forecasts than weekly. When compared to the
current forecast at HMS Networks, XGBoost and LightGBM performed better
based on both accuracy and bias. Overall, the findings from this study show that
machine learning can be applicable to demand forecasting when the data representation is aligned with the capabilities and limitations of the models. (Less)
Popular Abstract
A challenge that every company faces is the attempt to predict the future. This issue transcends most divisions, including sales and operations planning. Sadly, fortune tellers are of short supply. Thus, an array of methods to achieve accurate predictions have been developed over the years, including statistical models and machine learning algorithms.

The presence of machine learning (ML) tools in supply chain management has increased significantly over the past couple of years, with demand forecasting being the most common application. However, despite the common view on artificial intelligence as Mr Know-it-all, using ML to forecast demand has its limitations.
Asking an employee to predict the sales quantity for the next month will... (More)
A challenge that every company faces is the attempt to predict the future. This issue transcends most divisions, including sales and operations planning. Sadly, fortune tellers are of short supply. Thus, an array of methods to achieve accurate predictions have been developed over the years, including statistical models and machine learning algorithms.

The presence of machine learning (ML) tools in supply chain management has increased significantly over the past couple of years, with demand forecasting being the most common application. However, despite the common view on artificial intelligence as Mr Know-it-all, using ML to forecast demand has its limitations.
Asking an employee to predict the sales quantity for the next month will activate a reasoning that takes a large amount of factors into account. How much was sold last month? Do we have any new customers in the pipeline? Are there any issues in our production/delivery process? Is there a new competitor on the market? The questions go on and on. A highly skilled employee will deduce a reasonable prediction, but the risk of missing the mark remains.

Tree-based ML models, on the other hand, base their predictions solely on the historical information used to construct them. Most commonly, the historical sales data represents the main information source on which the models are trained. Thus, looking back to predict the future.

Using the rearview mirror creates an obvious vulnerability. What if our mirror is foggy? What if it is angled incorrectly? The forecasts are only as good as the data the models are trained on. The data available in practice offers an unlimited amount of flaws and inconcistencies. So, how do we handle imperfect data?

Picture a demand forecast through a magnifying glass. Is the demand forecasted on a monthly or weekly basis? Are we forecasting every unique article or larger product groups? When zoomed in, we get more precise information about when and how great the demand is, including the variance, but also the noise. On the other hand, zoomed out will look more stable and easier to predict, but result in less useful information for the demand planning operations.

The foundation of the thesis consisted of reviews of literature within demand forecasting and machine learning. Four ML models were developed and tested iteratively through design science research methodology. The findings in the thesis showed that the structure of the data as well as the data granularity had a larger impact on the results than the choice of model. Using methods such as first-order differencing and scaling allowed the ML models to generate forecasts that exceeded the performance of the current forecast at the case company.

However, the findings also showed the limitations of predicting through the rearview mirror. The questions posed through human intelligence offer useful insights. Insights that cannot be replaced by simple ML models looking back at historical data. Despite outperforming the alternatives, ML is not a fortune teller on its own. To create a fortune teller, one must therefore combine human intelligence and artificial intelligence. (Less)
Please use this url to cite or link to this publication:
author
Jhaveri, Jacob Filip LU and Skogman, Måns LU
supervisor
organization
course
MTTM05 20261
year
type
H2 - Master's Degree (Two Years)
subject
keywords
Demand Forecasting, Machine Learning, Supply Chain Management, Forecasting Metrics, Design Science Research
other publication id
6063
language
English
id
9240768
date added to LUP
2026-06-24 12:06:40
date last changed
2026-06-24 12:06:40
@misc{9240768,
  abstract     = {{This thesis investigates the applicability of using machine learning for demand
forecasting at HMS Networks. Using the design science research methodology, multiple forecasting artifacts were developed and tested iteratively, using
historical sales data. Four tree-based models; Decision Tree, Random Forest,
LightGBM and XGBoost, were tested with different pre-processing and feature
engineering techniques such as first-order differencing, scaling and varying levels of granularity. The results show that performance was primarily improved
through better data representations, rather than through model selection alone.
Particularly, first-order difference and scaling improved the models’ abilities to
generalize and reduced forecasting bias. Furthermore, the temporal granularity
was shown to strongly affect the stability of the forecast, where monthly granularity generated more stable forecasts than weekly. When compared to the
current forecast at HMS Networks, XGBoost and LightGBM performed better
based on both accuracy and bias. Overall, the findings from this study show that
machine learning can be applicable to demand forecasting when the data representation is aligned with the capabilities and limitations of the models.}},
  author       = {{Jhaveri, Jacob Filip and Skogman, Måns}},
  language     = {{eng}},
  note         = {{Student Paper}},
  title        = {{Machine Learning as a Tool for Demand Forecasting: A Study at HMS Networks}},
  year         = {{2026}},
}