The profitability of a supermarket is largely determined by efficient product display within promotional
bins and shelf displays. Hence it can be argued that a sustainable competitive advantage can be
achieved if demand of particular products could be accurately forecasted to allow for a customized
supply of consumer demand, ultimately leading to increased profitability. In this project, we focus on
one of the most basic and yet diversified products in terms of sizes within supermarkets – soda
beverages. Hence, our basic business objective is to develop a more targeted sales strategy of soda
bottles by leveraging the power of forecasting and predict the demand of soda bottles as to “when”
soda bottles of a particular size should be supplied for an optimized equilibrium of demand and supply.
Once defining the basic business objective, a first and particularly important step is an efficient data
preparation. From the initial data sheet, we aggregated the most commonly consumed soda beverages
to come up with four different container sizes, ranging in volume from 300ml to 2.25l. While preparing
the data for forecasting models, we paid particular attention to the power of visualization, which helped
us to improve and optimize our data and involved processes such as the identification of level, trend and
seasonality, the identification and exclusion of outliers caused by mystery bulk buyers, the choosing of a
weekly sales prediction, the focus on 2.25l sized containers and the exclusion of “noisy” weeks 49-57, all
leading towards an optimized data input allowing for accurate future predictions. It has to be noted that
in addition to the visualization and great effort the group put into the data preparation, external advice
was also gathered from a Indian based retail consultant, giving us further insight of how we can optimize
our data preparation to synergize our forecasting expertise with the initial business problem. Using this
data, a partitioning was conducted under best practice, creating a training and validation period which
was used for all forecasting model developments. The group saw the Holt-Winter’s to be the most fit in
order to capture the identified level, trend and possible seasonality but also ran a variety of other
models such as linear regressions, double exponential smoothing and a naïve forecast, serving as a
benchmark to compare the different measurements such as MAPE, MAD and MSE. Further, we also
analyzed possible autocorrelations of residuals and incorporated any significant ARs into our models if
found necessary. It turned out that the group’s initially used model, namely the additive Holt-Winter’s
model gave the lowest of all three measures above (64.7, 37.8 and 2332.2 respectively) and beat the
strong performance by merely using naïve forecasts from prior weeks. In other words, the improved
performance of our model can be effectively used by interested stakeholders of this Hypermarket: The
data shows that the model can predict future demand more accurately, leaving room for a wide range of
applications such as the establishment of a threshold that controls when 2.25l containers should be
displayed at certain promotional stands if a certain level is exceeded. However, it has to be noted that
despite the model has created more than positive results, additional data over time needs to be
gathered in order to further refine and improve the model to meet the dynamic demand of the 2.25l
soda size containers at Hypermarket and sustain this possible competitive advantage