Abstract:
This thesis joins the vivid details about pharmaceutical sales predictions using
predictions models, feature selections along with their respective analysis. Collectively,
existing thesis works from different researchers measures the accuracy using different
algorithms in different domains in terms of identifying ways to maximize the sales
using machine-learning processes. Distinctively, the process of finding the accurate
model by understanding the features that relatively affect for the accurate model will
be discussed throughout this paper.
This paper consists of factual pharmaceutical data in the area of Saskatchewan Canada,
which applied to the models discussed in the latter part of the study. As the
methodology for a successive implementation of accurate prediction
Comparing the model which were already implemented with time series analysis for
the retails, interpret their results with the previous studies and building the multivariate
time series model doing the feature engineering. External factors were taken into
consideration for the pharmaceutical sales while finding their effectiveness and feature
importance. Moreover, to analyse the correlation of the unit sales with other features
and provide the meaningful graph to visualize the data for the end user is an additional
important objective of the research.
Mainly this project concentrated on three time series models XGBoost, Regression,
ARIMA ,LSTM , to experiment pharmaceutical retail analysis. Analysing output result
of XGBoost, Regression model performance were thrived comparing to other models
by achieving 13% of MAPE while others achieved 22% and 53% of MAPE
respectively. This model has strength to predict the nearly 14 days of the future values,
hence this would be a beneficial for the decision makers to organizing their financial
plan and positioning the marketing campaigns effectively.
