Sales Forecasting Based on Economic Indicators for a Construction Company
Name
Andri Hõbemägi
Abstract
The construction sector's economic performance is closely related to macroeconomic conditions, shaping investment decisions and market dynamics. Despite its significance, forecasting sales within this sector is notably challenging due to the sector’s high sensitivity to economic fluctuations. Sales forecasting, in practice, traditionally relies on bottom-up approaches, where the prediction process starts from compiling individual business unit-level budgets and aggregates upwards to forecast the company's sales revenue. This method can overlook broader economic trends, failing to incorporate the vital interplay between market dynamics and company activities. Our study addresses this challenge by integrating both historical sales data and macroeconomic indicators into a unified forecasting model.
By constructing various feature sets through feature engineering and systematically evaluating them, we demonstrate how each set contributes to enhancing predictive performance. We applied five pre-selected regressors on each dataset, aiming to select the model with the lowest possible error with respective variability. As part of our pipeline, we also conducted optimisation of hyperparameters to fine-tune each regressor’s performance. We compared the results against the error threshold of 10%, which is of material importance as set by the Nasdaq Tallinn stock exchange for public companies to assess the relevance of the errors obtained from our models. This approach not only aligns company sales forecasts with external economic indicators but also refines the model's accuracy through targeted data enhancement and parameter optimisation. Conducted for Nordecon AS, a leading publicly listed construction company, this research provides a framework for robust quarterly forecasts that contribute to improving the company's decision-making processes.
By constructing various feature sets through feature engineering and systematically evaluating them, we demonstrate how each set contributes to enhancing predictive performance. We applied five pre-selected regressors on each dataset, aiming to select the model with the lowest possible error with respective variability. As part of our pipeline, we also conducted optimisation of hyperparameters to fine-tune each regressor’s performance. We compared the results against the error threshold of 10%, which is of material importance as set by the Nasdaq Tallinn stock exchange for public companies to assess the relevance of the errors obtained from our models. This approach not only aligns company sales forecasts with external economic indicators but also refines the model's accuracy through targeted data enhancement and parameter optimisation. Conducted for Nordecon AS, a leading publicly listed construction company, this research provides a framework for robust quarterly forecasts that contribute to improving the company's decision-making processes.
Graduation Thesis language
English
Graduation Thesis type
Master - Data Science
Supervisor(s)
Novin Shahroudi
Defence year
2024