Muhammad Wildan1, Moh Thaha Rizieq
Hentihu2, Riyanto Jayadi3
Universitas Bina Nusantara, Indonesia123
Email: [email protected]1,
[email protected]2, [email protected]3
|
Abstract |
|
PT.
XYZ is a retail company that has been operating since 2015 in New York, which
seeks to increase sales through attractive package bidding strategies and
ensuring stock availability for the following month's highest-selling items.
The project aims to analyze and forecast the items that will have the highest
sales as well as provide package recommendations that can increase sales. The
analysis was carried out using a machine learning model, with the Market
Basket Analysis (MBA) method for package recommendations and the Auto
Regressive Integrated Moving Average (ARIMA) for sales predictions. The
results of the analysis show that the "USB-C Charging Cable" and
"Bose SoundSport Headphones" have a strong association with the
"Vareebadd Phone", which indicates the
potential for increased sales if offered together. In addition, ARIMA's
predictions suggest that the MacBook Pro Laptop will generate the highest
revenue, with an average projection of around $788,067.45. Based on these
findings, we recommend PT. XYZ is preparing a package offering for the Vareebadd Phone with USB-C Charging Cable and Bose
SoundSport Headphones, as well as preparing stock for the MacBook Pro Laptop,
to meet market demand and maximize revenue potential. Keywords: Market Basket
Analysis, Auto Regressive Integrated Moving Average, Machine learning |
*Correspondence
Author: Muhammad Wildan
Email:
[email protected]
INTRODUCTION
PT.
XYZ, a retail entity founded in 2015 in New York, specializes in the sale of
electronic products through physical and online platforms, including Amazon. In
an effort to increase its sales figures, the company integrates package
offerings and aims to maintain a sufficient supply of products with high demand
every month (Datta et al., 2024;
Wu et al., 2022).� The novelty of this study lies in the
application of two different machine learning methodologies, namely Market
Basket Analysis (MBA) to recommend optimal product packages and Auto Regressive
Integrated Moving Average (ARIMA) to predict best-selling items. This approach
is expected not only to help PT. XYZ in proactively managing inventory, but it
also provides deeper insights into market dynamics that can improve their sales
and marketing strategies.
The
package offer policy provides an opportunity for consumers to acquire some
items at a lower price compared to individual purchases. Research by Chen et
al. (2016) shows that the implementation of bundling strategies can
significantly increase sales volume by utilizing existing product associations (D�Angelo &
Minchin, 2021). This is not only attractive
to discount seekers, but also to customers looking for complementary product
combinations (Wolle et al., 2019). In addition, the importance
of effective inventory management to maintain the availability of popular
products so as not to lose customers to competitors has also been discussed in
research by Huang and Sarig�ll� (2016) (Vermeulen et al., 2023).
To
achieve this, an efficient inventory management system is indispensable, which
can monitor stock levels in real-time and coordinate inventory across various
sales channels, both online and offline. Identifying potential items for
bundling requires an in-depth understanding of the relationships between
various products (Priessner &
Hampl, 2020; Volles et al., 2024). Association analysis plays
an important role in this context, as it explores the relationships between
items in a dataset. Techniques such as the A priori algorithm can be used to
identify sets of items that often occur together in transactions above a certain
threshold. By identifying these associations, researchers can strategically
place related products to increase sales (Martinez et al., 2021).
Additionally,
to understand patterns and predict future trends, predictive analytics is used,
often utilizing machine learning techniques to improve accuracy and insights in
data handling, pattern detection, and outcome prediction (Dubey et al., 2019;
Gunasekaran et al., 2017). Previous research has shown
that market analysis has been widely used by various companies to inform their
promotional strategies based on identified product associations (Qisman et al., 2021). For example, Limitedbrands
has shown how targeted promotional campaigns, such as "buy two, get
three" or "buy products, get rewards", can greatly benefit from
insights gained from market analysis by ensuring that the promoted items are
paired appropriately (Kurnia et al., 2019).
To
forecast inventory needs and sales trends, various advanced statistical and
machine learning models are applied. These include the Fuzzy Logic (FL)
approach to optimize inventory order quantities (Li et al., 2020), Artificial Neural Network
(ANN) to predict stock market trends, and the Autoregressive Integrated Moving
Average (ARIMA) method to forecast sales and stock prices (Aldino et al., 2021). In this study, we involved
two different machine learning methodologies: Market Basket Analysis (MBA) to
recommend optimal product packages and Auto Regressive Integrated Moving
Average (ARIMA) to predict best-selling items, thus allowing PT. XYZ to
proactively manage inventory in order to meet anticipated demand. By combining
these two approaches, this research provides a strong foundation for PT. XYZ to
improve their sales and inventory management strategies, as well as strengthen
their competitive position in the electronics retail market.
RESEARCH METHODS
Data Understanding
This data comprises sales
transaction data from January to December 2019, totaling 185,919 transactions.
Among these transactions, 7,136 indicate more than 1 product being bought
together (Monteserin &
Armentano, 2018).
Table 1. Data Understanding Table
|
Attribute name |
Data type |
Description |
|
OrderID |
Numerical - Discrete |
Unique identifier
assigned to each individual order or purchase made by a customer. |
|
Product |
Categorical - Nominal |
The name or description of the item(s) that
were purchased in the order |
|
Quantity |
Numerical - Discrete |
The number
of units or items of a particular product that were
purchased in the order. |
|
Price |
Numerical - Discrete |
The cost
or price of a single
unit of the product |
|
OrderDate |
Categorical - Ordinal |
The date
and time when
the order was placed by the customer |
|
Purchase Address |
Categorical �
Nominal |
The location where
the order was delivered or the customer's billing address |
Data Preparation
1) Merging data
The data we get is sales data
per month for 1 year in 2019, therefore we first merge the data before
continuing with the next data preparation process and also import data into
RapidMiner.
Fig 1. Merge Data
2) Delete Empty Row
In the data we get, after
several rows of data, there will usually be a row whose contents are empty
before moving on to the next row, and this happens consistently across all
sheets. Because of this, we delete this empty row so that the data remains
sorted from above without being truncated.
Fig 2. Empty Row
3) Reformat Order Dat
In the data we got, there were
inconsistent date data formats, some were automatically converted to date, but
some were still strings. This is one of the factors that makes it difficult to
enter existing data into rapidminer. Therefore, we need to change the date
format to be consistent with the same date format.
Fig 3. Reformat Order Date
Modelling
a. Market Basket Analysis
The Market Basket Analysis
(MBA) model has been made using rapidminer. In order to make the data usable in
rapidminer, the data that has been preprocessed by merging and performing
necessary data cleaning steps needs to be processed again using rapidminer. We
first import our data into rapidminer and start the data manipulation using
rapidminer operators.
Fig 4. Market Basket Model
Our goal here is to make our
data set readable by Rapidminer and to do that the operators we used are
Aggregate, Pivot, Set Role, Rename by Replacing, Replacing Missing Values, and
Numerical to Binomial. On the aggregation attributes we choose Quantity Ordered
with sum functions, and we group it by Order ID and Product. and by using Pivot
we group the attribute by Order ID and column grouping attribute using Product
also setting the aggregation attributes by Quantity Ordered. We set the role of
Order ID to id and simplify the name of each product using the Rename By
Replacing operator. We replace the missing value with zero and change the
numerical value to binomial. From this data preparation the output that we want
is to place the Order ID on the left column and placing all the Product Name on
the top row. Each column under the Product Name will be true if the product is
purchased from the Order ID on the left column, otherwise it will be false. The
output will be shown using the Apply Association Rules operator.
Fig 5. Market Basket Model Extend
The model for Market Basket
Analysis can be seen on the figure. The data that has been formatted is placed
on the Data Prep operator. Here we apply FP-Growth and Association Rules in our
model. By implementing FP-Growth our aim here is to construct an FP-Tree and
mine frequent items. We also apply Association rules to find the important
parameters on making decisions for bundling which are Support, Confidence, and
lift. Support is the percentage of transactions that contain all of the items
in an itemset. High support values indicate that the items in the itemset are
often bought together, making them good candidates for bundling. Confidence is
the probability that a transaction that contains the items on the left hand
side of the rule also contains the item on the right hand side. Higher
confidence values suggest a stronger likelihood that the presence of itemset A
will be followed with the presence of itemset B.
Lift is the probability of all
of the items in a rule occurring together divided by the product of the
probabilities of the items on the left and right hand side occurring as if
there was no association between them. If the lift is greater than 1, it
implies a positive correlation between items A and B. This indicates that the
presence of itemset A in a transaction increases the likelihood of itemset B
being present. If the lift is equal to 1, it indicates that there's no
association between the two itemsets. If the lift is less than 1, it implies
that customers who buy itemset A are less likely to buy itemset B. After
importing the data and doing the required data reprocessing, we divide the data
using the Split Data operator into FP-Growth and Apply Association Rules
operator. On our model the parameter on FP-Growth is set on default and
connected to the Create Association Rules operator. We set the criterion in
Create Association Rules operator to confidence and set the confidence to 0.001
to get more results. Finally, we use the Multiply operator to copy the format
from Create Association Rules to output and into the Apply Association Rules
operator. In order to get the result from FP-Growth we connect the item sets of
the Create Association Rules operator to the output. Finally, by giving the
Apply Association rules operator the required data we connect the example set
into the output. This ends us with 3 types of output which are Frequent item
sets (FP-Growth), Association Rules (Create Association Rules), and Example
sets (Apply Association Rules).
b. ARIMA (auto Regressive
Integrated Moving Average)
The ARIMA model is performed
using Python. Before performing modelling, the dataset will require additional
preparation by splitting data into training and testing sets (Kaur & Kang,
2016). The process of splitting the
dataset will be based on the date and time marker on each transaction that
happened. Below, is showed the code on how the data will be split into two
different sets based on the date.
Fig 6. Arima Model
In this process, the training
dataset is all the transactions of every product that happened from January
1st, 2019, until November 30th, 2019. For the testing dataset, it will be
filled with all transactions of every product from December 1st, 2019, until
December 31st, 2019.
Fig 7. Arima Model Extend
After split the datasets we
will be using method ARIMA that we can just import from statmodels package
(statsmodels.tsa.arima.model). The ARIMA will have 3 important parameters: p is
the number of autoregressive terms, d is the number of nonseasonal differences
needed for stationarity, and q is the number of lagged forecast errors in the
prediction equation. For the initial training, we will use p = 5, d = 1, and q
= 2 as they are known for the best number for the ARIMA setup. Then we will do
model.fit() to fit the training dataset into the model with set parameters.
Then, we can do testing by using model_fit.predict into the testing dataset so
that we are able to know how the model is performed.
RESULTS
AND DISCUSSION
Market Basket Analysis
Fig 8. Market Analysis Result
The figure shows the result of
the data that shows the products and products bought together with the highest
support by using the FP-Growth operator in Rapidminer. From this data, it can
be seen that the product most purchased is Wired Headphones with the support of
0.10596 and most purchased products together with confidence of 0.00558, which
consist of USB-C Charging Cable and Google Phone. Based on this information, we
can see that the item bought together is connected in terms of being used
together. This will help us in making the decision on which products should be bundled
together.
Fig 9. Association rules
The result we got from the
Create Association Rules operator shows the relation of each possible item
combination. This result also shows all required parameters to make the
important decision for the bundling of the product which is Support,
Confidence, and Lift. We sort the result table to show the highest possible
Support, Confidence, and Lift to decide the best possible items to bundle
depending on the purpose of the bundle. Initial inspection of the dataset via
the Create Association Rules operator shows recurring item sets, with the
standout being the "Google Phone" and "USB-C Charging
Cable" pairing, boasting the highest support value of 0.00558. This
finding indicated a strong tendency for these items to be purchased together,
prompting consideration of a bundling strategy to enhance sales.
Adding to this, the analysis
shows the item set with the highest confidence with a notable 0.50549,
connecting "Lightning Charging Cable," "Wired Headphones,"
and "iPhone." This high confidence underscored a robust association,
implying that customers purchasing the latter item also consistently opted for
the former, reinforcing cross-selling and promotional opportunities. Furthermore,
the highest lift value of 15.47055 shows a connection for "USB-C Charging
Cable" and "Bose SoundSport Headphones" combination in pair with
the "Vareebadd Phone". With this, the idea of lift goes beyond just
showing connections. It suggests that when you buy certain items together, it
makes it much more likely that you'll also end up getting the "Vareebadd
Phone". Discovering the highest support, confidence, or lift values for
itemsets unveils significant purchase patterns. High support signifies
frequently co-purchased items, guiding inventory decisions. High confidence
highlights strong item associations, informing cross-selling approaches. A high
lift value indicates meaningful item connections beyond chance, aiding bundling
and promotional strategies. Ultimately, these insights empower businesses to
optimize inventory, enhance cross-selling, and strategically bundle items for
improved customer experience and revenue growth.
Based on our analysis of the
association rules, we have identified several promising opportunities for
bundling products. The itemset with the highest support, consisting of the
"Google Phone" and "USB-C Charging Cable," presents an
ideal pairing that could be marketed together as a bundle, potentially boosting
sales for both items. Additionally, we observed an intriguing relationship
involving the "USB-C Charging Cable," "Bose SoundSport
Headphones," and "Vareebadd Phone." Specifically, when customers
purchase the charging cable and the headphones, they are likely to also buy the
Vareebadd Phone. This insight suggests a strategic opportunity for Willbert
Electronics to enhance product placement by positioning these items in close
proximity to one another or creating a bundled offer that includes all three
products.
The impact of such bundling
strategies on business performance could be significant. By leveraging the
identified associations, Willbert Electronics can increase the average
transaction value and improve customer satisfaction through convenience, as
customers appreciate the ease of purchasing complementary products together.
Furthermore, these bundles can attract price-sensitive consumers who are
looking for deals, thereby expanding the customer base. From a stock management
perspective, accurately predicting the demand for bundled items can lead to
more effective inventory management, ensuring that popular products are readily
available and reducing the risk of stockouts. This proactive approach to
inventory forecasting not only enhances sales but also strengthens customer
loyalty, as satisfied customers are more likely to return for future purchases.
Regression Analysis
From the modelling, we can
predict all of the product selling performance within the month of December.
Here we will be using a green line as a marker for the prediction result for
the next 30 days after November 2019. In this analysis, we will only show the
top 5 products with the highest selling dollar based in the month of December.
Below are showed 5 plots showing the blue line which was the data used to train
the model, the orange line that is used for testing the model, and the green
line is the predicted revenue.
Fig 10. Arima Result
The plots above are showed to
be the top 5 highest products that create revenue the most. In details, the
table below showing the predicted revenue in the month of December for the top
5 products, by total amount, and by the daily average
Table 2. Highest Product
|
Total
Predicted Revenue |
Daily
Average Predicted Revenue |
|
|
Macbook Pro Laptop |
788,067.45 |
25,421.53 |
|
iPhone |
488,233.72 |
15,749.47 |
|
ThinkPad Laptop |
363,486.46 |
11,725.37 |
|
Google Phone |
305,502.62 |
9,854.92 |
|
27in 4K Gaming
Monitor |
255,158.02 |
8,230.90 |
The ARIMA model evaluated by
using Root Mean Square Error (RMSE) and Mean Absolute Percentage Error (MAPE).
Both metrics used for standard measurement for regression analysis. The RMSE
can tell how many units the model is wrong on average. While MAPE tells how
wrong the forecasts are percentage- wise, for example if the MAPE value 0.02 it
means that the forecast are 98% accurate.
The overall error metrics for
the ARIMA models across all products are as follows:
a. Overall Root Mean Square Error
(RMSE): 4049.61
b. Overall Mean Absolute
Percentage Error (MAPE): 0.288 or 28.8
The RMSE of 4049.61 indicates
that, on average, the model's predictions are around 4049.61 units away from
the actual values. The MAPE of 28.8 suggests that the model's predictions are
off by about 28.8%. Based on the predictions result above, related to the
stocks, it is advisable for the owner to bundle product that includes the one
of the or both form the 5 products that dominate the revenue the most.
CONCLUSION
Exploring
the realm of Market Basket Analysis (MBA) for bundling in an electronics store environment
presents several promising avenues for future development. Considering the
dynamic nature of electronics and rapid technological advancements, integrating
product lifecycle data could unveil insights into optimal bundling strategies
that align with product trends and obsolescence. Additionally, incorporating
external factors such as customer reviews, ratings, and online discussions
related to electronic products could enrich the analysis, while integrating
sentiment analysis would provide a deeper understanding of consumer preferences
and the effectiveness of the suggested bundling strategy. Currently, the ARIMA
model produces a flat prediction line, indicating it forecasts the same value
for all future data points; to address this, transforming data to stabilize
variance and decomposing seasonal components could enhance accuracy. Feature
engineering techniques, such as using lagged variables and incorporating
external variables as exogenous factors, may also yield better predictions. The
implications of this research extend beyond sales optimization, as refining
bundling strategies and enhancing inventory forecasting can improve customer
satisfaction and loyalty, ultimately driving long-term profitability.
Furthermore, integrating customer sentiment analysis can lead to more targeted
marketing efforts, aligning offerings with consumer desires, while future
studies might explore the impact of promotional campaigns on bundling
effectiveness, contributing valuable insights to the broader field of retail
analytics.
Aldino, A. A., Pratiwi, E. D., Setiawansyah, Sintaro, S.,
& Putra, A. D. (2021). Comparison of Market Basket Analysis to Determine
Consumer Purchasing Patterns Using Fp-Growth and Apriori Algorithm. 2021
International Conference on Computer Science, Information Technology, and
Electrical Engineering, ICOMITEE 2021.
https://doi.org/10.1109/ICOMITEE53461.2021.9650317
D�Angelo, D., & Minchin, R. E. (2021). Project bundling
in transportation construction. Proceedings of International Structural
Engineering and Construction, 8(1).
https://doi.org/10.14455/ISEC.2021.8(1).CON-09
Datta, A., Sarkar, B., Dey, B. K., Sangal, I., Yang, L., Fan,
S.-K. S., Sardar, S. K., & Thangavelu, L. (2024). The impact of sales
effort on a dual-channel dynamical system under a price-sensitive stochastic
demand. Journal of Retailing and Consumer Services, 76, 103561.
https://doi.org/10.1016/j.jretconser.2023.103561
Dubey, R., Gunasekaran, A., Childe, S. J., Papadopoulos, T.,
Luo, Z., Wamba, S. F., & Roubaud, D. (2019). Can big data and predictive
analytics improve social and environmental sustainability? Technological
Forecasting and Social Change, 144, 534�545.
https://doi.org/10.1016/j.techfore.2017.06.020
Gunasekaran, A., Papadopoulos, T., Dubey, R., Wamba, S. F.,
Childe, S. J., Hazen, B., & Akter, S. (2017). Big data and predictive
analytics for supply chain and organizational performance. Journal of
Business Research, 70, 308�317.
https://doi.org/10.1016/j.jbusres.2016.08.004
Kaur, M., & Kang, S. (2016). Market Basket Analysis:
Identify the Changing Trends of Market Data Using Association Rule Mining. Procedia
Computer Science, 85. https://doi.org/10.1016/j.procs.2016.05.180
Kurnia, Y., Isharianto, Y., Giap, Y. C., Hermawan, A., &
Riki. (2019). Study of application of data mining market basket analysis for
knowing sales pattern (association of items) at the O! Fish restaurant using
apriori algorithm. Journal of Physics: Conference Series, 1175(1).
https://doi.org/10.1088/1742-6596/1175/1/012047
Li, H., Wu, Y. J., & Chen, Y. (2020). Time is money:
Dynamic-model-based time series data-mining for correlation analysis of
commodity sales. Journal of Computational and Applied Mathematics, 370.
https://doi.org/10.1016/j.cam.2019.112659
Martinez, M., Escobar, B., Garc�a-D�az, M. E., &
Pinto-Roa, D. P. (2021). Market basket analysis with association rules in the
retail sector using Orange. Case study: Appliances sales company. CLEI
Eletronic Journal (CLEIej), 24(2).
https://doi.org/10.19153/cleiej.24.2.12
Monteserin, A., & Armentano, M. G. (2018).
Influence-based approach to market basket analysis. Information Systems,
78. https://doi.org/10.1016/j.is.2018.01.008
Priessner, A., & Hampl, N. (2020). Can product bundling
increase the joint adoption of electric vehicles, solar panels and battery
storage? Explorative evidence from a choice-based conjoint study in Austria. Ecological
Economics, 167, 106381.
https://doi.org/10.1016/j.ecolecon.2019.106381
Qisman, M., Rosadi, R., & Abdullah, A. S. (2021). Market
basket analysis using apriori algorithm to find consumer patterns in buying
goods through transaction data (case study of Mizan computer retail stores). Journal
of Physics: Conference Series, 1722(1).
https://doi.org/10.1088/1742-6596/1722/1/012020
Vermeulen, H., Meyer, F., & Sch�nfeldt, H. C. (2023). A
basic healthy food basket approach to evaluate the affordability of healthy
eating in South Africa and Kenya. Frontiers in Sustainable Food Systems,
7. https://doi.org/10.3389/fsufs.2023.1181683
Volles, B. K., Ribbers, D., Van Kerckhove, A., & Geuens,
M. (2024). Beyond bundles: Choosing product bundles increases shopping basket
size. Journal of Retailing and Consumer Services, 81, 104035.
https://doi.org/10.1016/j.jretconser.2024.104035
Wolle, M. M., Stadig, S., & Conklin, S. D. (2019). Market
basket survey of arsenic species in the top ten most consumed seafoods in the
United States. Journal of Agricultural and Food Chemistry, 67(29).
https://doi.org/10.1021/acs.jafc.9b02314
Wu, X., Zha, Y., & Yu, Y. (2022). Asymmetric retailers�
sales effort competition in the presence of a manufacturer�s help. Transportation
Research Part E: Logistics and Transportation Review, 159, 102625.
https://doi.org/10.1016/j.tre.2022.102625
|
|
� 2025 by the authors. Submitted for possible open access publication
under the terms and conditions of the Creative Commons Attribution (CC BY SA) license (https://creativecommons.org/licenses/by-sa/4.0/). |