1. Application of machine
learning techniques for
sales forecasting
Jinxing LIN
Supervisor: Dr Irene Moulitsas
School of Engineering
Cranfield University
Background Key Results
ConclusionResearch
MSc Computational and Software Techniques in Engineering YYYY (????? Option)
As a matter of fact that the dataset we work
with is a time-series dataset, the project
mainly focus on exploring time-series ML
algorithms such as exponential smoothing,
auto-regressive integrated moving average
(ARIMA).
Machine learning (ML) has been the hottest
topic of Information Technology (IT) during this
decade. It mainly uses algorithms to explore
data, determine relationships among data and
make predictions. Machine learning
algorithms are applied to solve a lot of
forecasting problems.
Sales forecasting is one of these problems.
We aim at discovering different ML techniques
and using them to train with a real dataset in
order to make some relatively accurate
predictions for the future sales.
A precise sales forecasting helps a company
to better manage its resources and therefore
leads to a larger benefit.
Due to the irregularity of data and the limited
historical sales data of some products, some
typical time-series training algorithms don't
perform well with the data in this dataset. The
customised models outperform the others
including the forecasting model currently used
by the company Didactic. It gives a smaller
error which is 2.2% smaller comparing to
Didactic's model.
●
Study machine learning algorithms;
●
Prepare and pre process data;
●
Visualise and understand data;
●
Analyse data before building models;
●
Apply typical time-series machine learning
techniques;
●
Design and implement several customised
forecasting models specifically for this
dataset;
●
Compare and analyse results.
x
On the other hand,
some customised
models are
constructed for
improving the
accuracy of the
forecasting results. These customised models
are all half individual half common models:
●
a trend built with regression algorithms for
each product (individual);
●
a common set of seasonal adjustments for all
products (common).
Further work
●
Apply neural network for obtaining a better
regression for the calculation of trend;
●
Integrate clusters algorithms into the model.
Fig1 : Time-series model
Fig 2 : Customised model training process
Fig 3 : Comparison of global cost
Fig 4 : Historical sales and forecasting – ref 1
Fig 5 : Historical sales and forecasting - ref 2