International Journal of Data Mining & Knowledge Management Process (IJDKP), Vol.13, No.

4/5, September 2023

WHEN SHOULD THEY BUY?

SURVIVING MACHINE LEARNING MODELS FOR
PURCHASE TIMING
Diego Vallarino

Independent Researcher, Madrid, Spain.

ABSTRACT
Unlocking the value of raw data requires transforming it into knowledge and information that drives
decision-making. Algorithms for Machine Learning (ML) are capable of analyzing big datasets and
delivering accurate predictions. Using machine learning, market segmentation, customer lifetime value,
and marketing approaches have been used.

This article explores machine learning approaches for marketing, including Support Vector Machines,
Genetic Algorithms, Deep Learning, and K-Means. It is rare for ML to be used to forecast when a person
will purchase a product or a basket of goods.

The survival models Kernel SVM, DeepSurv, Survival Random Forest, and MTLR are evaluated to predict
individual tine-purchase choices in this research. Gender, Revenue, Location, Purchase History, Online
Behaviour, Interests, and Promotions According to the research, both discounts and CustomerExperience
have an impact on the amount of time spent making a purchase. The research indicates that the DeepSurv
model best predicted purchase completion. These insights aid in raising conversion rates for marketers.

KEYWORDS
Survival Analysis, Machine Learning for survival, marketing analytics.

1. INTRODUCTION
Due to advances in information technology and the meteoric emergence of the Internet, the data
revolution of the last few decades has forced organizations to generate more information than
they can effectively employ or comprehend (see Erevelles, S.; Fukawa, N.; Swayne, L., 2016;
Seng, J.L.; Chen, T., 2010). The development of data volume, data diversity, and analytical
breadth has prompted technology breakthroughs beyond storage, transit, and processing (see
Seng, J.L.; Chen, T., 2010)

The data must be converted into knowledge and information in order to convey knowledge into
corporate decision-making tools. This information is used in marketing research to establish
interesting connections between market segmentation in industrial, tourist, and other markets,
customer lifetime value, loyalty and client segment, direct market, marketing campaign, and
other applications (Tkác, M.; Verner, 2016). Using Machine Learning (ML) techniques (see
Bahari, T.F., &Elayidom, M.S., 2015; Jessen, H.C., &Paliouras, G., 2001), it is envisaged that
huge quantities of stored data may be examined, and useful information retrieved.

ML are strategies that equip computers with the capacity to comprehend, using data and
experiences similar to the human brain (Çelik, Ö., 2018). The application of ML models to data is
DOI:10.5121/ijdkp.2023.13501 1
 International Journal of Data Mining & Knowledge Management Process (IJDKP), Vol.13, No.4/5, September 2023

motivated by the fact that these models can be applied to the resolution of a variety of problems,
ranging from those that could be solved through conventional statistics and management of
scientific techniques to complex problems requiring a larger analysis; this is because ML allows
solving problems more quickly and effectively than conventional methods. These strategies are
used to forecast the outcome of new data, make predictions, classify information, or aid in the
decision-making process.

Although many academics have worked with ML approaches in marketing, it is vital to

summarize the numerous ways that may be used to certain areas, as well as how these methods
might maximize outcomes and contribute to marketing-related firm decision making. Important
to the study of the future of these techniques is the management of data; therefore, a review of the
different methods commonly used in marketing has been conducted, with the goal of generating
technologies and resolution of problems in a short period of time, with excellent results, such as
Support Vector Machine (SVM), Genetic Algorithms (GA), Deep Learning, and K-Means, among
others.

Several researchers have investigated marketing techniques. In the current state of the art, (see
Sun, Z.; Guo, Q.; Yang, J.; Fang, H.; Guo, G.; Zhang, J.; Burke, R., 2019)describes the
recommendation system's algorithms. They demonstrate the progression of approaches used to
generate suggestions, such as collaborative algorithms for filtering based on proximity and
model-based approximations.

The use of Artificial Neural Networks (ANN) for regression, analysis of primary components,
and classification as a tool to estimate consumer behavior is described in (Briesch, R.; Rajagopal,
P., 2018). They discuss the benefits and drawbacks of their usage of ANN in marketing and the
application according to the evaluation of the disclosed by other authors in essence, behavior,
satisfaction, customer segmentation, or the selection of a new product or response to publicity.
(Chen, S.H.. 2016), illustrates the several methods that may be used to anticipate the departure of
customers from various sorts of businesses(see Liu, X.; Singh, P.V.; Srinivasan, K., 2016).

These algorithms consist of, among others, support vector machines (SVM), decision trees (DT),
and artificial neural networks (ANN). To conduct their study, authors such as (Balducci, B.;
Marinova, D., 2018;Liu, X.; Singh, P.V.; Srinivasan, K., 2016)employ unstructured data. To
estimate consumer demand for TV series, acquire data from internet sources such as Twitter
Tweets, Google Trends, Wikipedia visits, IMDB ratings, and Huffington Post articles (see Liu,
X.; Singh, P.V.; Srinivasan, K., 2016). To analyze a large quantity of unstructured data, these
writers use a variety of machine learning algorithms, text mining, and cloud computing, among
others.

They observed that there is no clear correlation between the number of Twitter followers and the
number of reviews acquired, projected the most-watched episodes, and concluded that this might
have a future impact on demand, demographic distribution, and other elements that drive
consumption. (Balducci, B.; Marinova, D., 2018), conduct a thorough analysis of the many
approaches and software that may be used to handle unstructured data, and present a complete
work framework to show the theoretical depth and computer advancements in marketing.
Regarding customer decisions, (Hauser, J.R., 2014; Dzyabura, D.; Hauser, J.R., 2011)study the
application of machine learning to the evaluation of the consumer's heuristics for selecting a
series of items, employing a set of decision rules.

In Duarte, V., et al. (2018),it shows that marketing applications and ML approaches increased.
Scientific productivity grew, but mainly in low-quality journals (Q2 and Q3) without the bottom
quartile (Q4). Quotes decreased with time.ML approaches vary. In this era, deep learning was the

2
 International Journal of Data Mining & Knowledge Management Process (IJDKP), Vol.13, No.4/5, September 2023

most utilized approach, but in 2019, it became one of many. Due to greater technique-marketing
subarea specificity, this is possible. They found that until 2008, marketing apps employed some
tactics broadly, but in 2019, they became more targeted.

As a new approach, survival analyses focus on the time of the event of interest, complementing
studies that examine whether a consumer will buy or act in a given manner. Survival analysis
seldom predicts purchases. Time-varying variables and censoring in modeling make survival
analysis better than static classification approaches.

The Cox proportional hazard model was initially used to predict bank failure for example, like
Lane, Looney, and Wansley (1986) showed. Luoma and Laitinen (1991) used Cox proportional
hazard models to predict Finnish industrial and retail firm failure, but they were inferior to
discriminant and logit analysis. Shumway (2001) models discrete-time bankruptcy hazard using
accounting and market data. Chava and Jarrow (2004), Carling, Pan, Ariyan, Narayan, and Truini
(2007), Leong, Nguyen, Meredith, et al. (2008), and Leonardis & Rocci (2009) employed the
discrete hazard model because of its parameter calculation advantages and the sort of variables
organizations commonly submit (2008).

Logistic regression and Cox proportional hazards regression disregard repeating occurrences or
fail to account for within-subject correlation, resulting in inaccurate standard error estimates and
a departure from the original study issue (Twisk, Smidt, & de Vente, 2005). Many methods have
been developed to analyze recurrent events using all available data and within-subject
correlations. Marginal intensity approaches, based on various risk set definitions, allow all cases
to be at risk for each repeated event (Wei, Lin, &Weissfeld, 1989), while conditional intensity
models are estimated in elapsed time or gap time and only designate cases at risk for the kth
repeated event after the (k-1)th event (Andersen & Gill, 1982; Chang & Wang, 1999; Prentice,
Williams, & Peterson, 1981).

In the survey of the literature that we have carried out for this work, there is no relevant material
that studies the time until purchase, using machine learning models under the survival
methodology. As a consequence, we may conclude that survival analysis is not the major focus of
specialists in purchase prediction. Our study seeks to evaluate the applicability of survival
analysis (SA) to purchase decision.

In this study, we investigate the outcomes of our model comparison and their economic
interpretation. Our investigation focuses on the ability of several models to forecast time-to-
purchase based on a collection of pertinent data. Specifically, we examine the predictive power of
the Kernel SVM, DeepSurv, Survival Random Forest, and MTLR survival models. We utilize the
concordance index to compare the various machine learning techniques (C-index)

Our objective is to determine which model delivers the most accurate and useful forecasts of the
time of purchase and to understand the economic relevance of the model's findings. To do this,
we evaluate the relevance and size of the predicted coefficients for each model variable and
compare these findings to economic theory and common sense. By studying the outcomes of our
model comparison and the economic interpretation of these results, we seek to gain insight into
the elements that lead to buy moment and a better knowledge of how various models may be used
to anticipate these specific occurrences.

The remaining sections of the paper are organized as follows. Following is a description of the
empirical analysis portion, including the used models, data source, and assessment measures.
After presenting the results of the investigation, including a comparison of the various models,

3
 International Journal of Data Mining & Knowledge Management Process (IJDKP), Vol.13, No.4/5, September 2023

we examine the economic implications of our findings. Finally, we summarize the important
results and their implications for future study and policymaking to end the work.

Overall, our work adds to the literature on the use of survival analysis in finance and sheds light
on the causes that lead to financial collapses, therefore assisting policymakers in designing more
effective rules to avoid similar disasters in the future.

2. EMPIRICAL ANALYSIS
In this section, we present the empirical analysis of the purchase moment. We analyze set of
information from real-online customers, who have decided to buy or not to buy a specific basket
of products, in a period of 12 months. The dataset consists of 10 variables, each of which
contributes to understanding the factors influencing the timing of a purchase. The study of
purchase moment is of great importance in marketing and behavioral economics, as it has
significant implications retailer knowledge. In this section, we describe the models used in our
analysis, the data source, and the evaluation metrics.

2.1. Models

2.1.1. Cox Proportional Hazards Model (coxph)

The Cox proportional hazards model is a widely used semi-parametric model in survival analysis.
It assumes that the hazard function can be represented as the product of a time-independent
baseline hazard function and a time-varying covariate function. Mathematically, the model can be
represented as:

ℎ(𝑡|𝑥) = ℎ0 (𝑡) exp(β𝑇 𝑥)

where ℎ (𝑡|𝑥 ) is the hazard function for a given time t and covariate values 𝑥, ℎ0(𝑡) is the
baseline hazard function, 𝛽 is a vector of regression coefficients, and 𝑒𝑥𝑝(𝛽𝑋) is the hazard ratio,
which represents the change in hazard associated with a unit change in the covariate.

2.1.2. Multi-Task Logistic Regression (MTLR)

Multi-task logistic regression is a machine learning method that can be used for survival analysis.
It is a multi-output learning algorithm that can predict the probability of an event occurring at
different time points. Mathematically, the model can be represented as:

𝐾 𝑝
ℎ(𝑡|𝑥) = 𝑒𝑥𝑝(𝛴𝑘=1 𝛴𝑗=1 𝛽𝑘𝑗 𝑥𝑘𝑗 )

Where ℎ(𝑡|𝑥) is the hazard rate for an individual with covariates x , βkj are the regression
coefficients for the kth characteristic of the jth group, and 𝑥kj is the kth feature of the jth group.

2.1.3. Random Survival Forest

Random survival forests are an extension of random forests for survival analysis. They use an
ensemble of decision trees to predict the survival function. The model can be represented as:

ℎ(𝑡|𝑥) = (1/𝐵)ΣBb=1 ℎ𝑏 (𝑡|𝑥 )

4
 International Journal of Data Mining & Knowledge Management Process (IJDKP), Vol.13, No.4/5, September 2023

Where ℎ𝑏 (𝑡|𝑥 ) is the hazard rate for an individual with covariates 𝑥 in the 𝑏𝑡ℎ decision tree and
𝐵 is the number of trees in the random forest.

2.1.4. DeepSurv

DeepSurv is a deep learning model for survival analysis. It uses a neural network with a flexible
architecture to predict the survival function. The model can be represented as:

𝑝
ℎ (𝑡|𝑥 ) = 𝑒𝑥𝑝 (𝛴𝑖=1 𝛽𝑖 𝑓𝑖 (𝑥) + 𝑔(ℎ𝜃 (𝑥)))

Where ℎ(𝑡|𝑥) is the hazard rate for an individual with covariates 𝑥$, $β_𝑖 are the regression
coefficients for the input features 𝑓𝑖 (𝑥), 𝑔(⋅) is a non-linear function that transforms the output
features and ℎθ (𝑥) is a neural network with θ parameters.

2.2. Data

The dataset with which we have worked is a set of information from real-online customers, who
have decided to buy or not to buy a specific basket of products, in a period of 12 months. The
dataset consists of 1000 different people, with 10 variables, each of which contributes to
understanding the factors influencing the timing of a purchase. Forty-eight percent of data
censored by the right, which implies that the rest of the customers did buy the basket of products
that was the focus of the investigation.

It contains information necessary for conducting an analysis on the time until a purchase of a
specific basket of products. In this analysis, we will explore the components of each variable and
examine their potential impact on the time to purchase.

2.2.1. Variable Components

Age: This variable represents the age of the individuals. Age can be a significant factor in
determining purchasing behavior, as consumer preferences and needs often vary across different
age groups. Gender: The gender variable indicates the gender of the individuals. Gender can play
a role in shaping consumer preferences and purchase decisions, as it may influence product
preferences, brand loyalty, and shopping behavior.

Income: The income variable reflects the income levels of the individuals. Income is an essential
determinant of purchasing power and can influence the affordability and willingness to make a
purchase. Marital Status: This variable captures the marital status of the individuals. Marital
status can affect purchasing decisions, as the needs, priorities, and spending patterns of
individuals may differ based on their marital status.

Location: The location variable represents the geographic location of the individuals.
Geographical factors such as culture, availability of products, and local market dynamics can
influence purchasing behavior. Purchase History: This variable indicates the individual's past
purchase history. Previous purchasing behavior can be an indicator of future purchase intentions,
as individuals with a higher purchase history may exhibit greater brand loyalty or a higher
propensity to make repeat purchases.

Online Behavior: The online behavior variable reflects the level of engagement of individuals in
online activities related to shopping. Online behavior can influence the time to purchase, as
individuals who are more active online may have access to a wider range of products,

5
 International Journal of Data Mining & Knowledge Management Process (IJDKP), Vol.13, No.4/5, September 2023

information, and promotions. Interests: This variable captures the specific interests of the
individuals. Individual interests can influence purchase decisions, as individuals are more likely
to purchase products or services related to their specific areas of interest.

Promotions and Discounts: The promotions and discounts variable indicate whether individuals
are exposed to promotional offers and discounts. Promotional activities can influence the time to
purchase by creating incentives for individuals to make a purchase sooner or by influencing their
decision-making process. Customer Experience: This variable represents the individuals' past
experiences with customer service. Customer experience can impact the time to purchase, as
positive experiences may enhance brand perception and increase the likelihood of repeat
purchases.

By analyzing these variables and their components, we aim to gain insights into the factors that
contribute to the timing of a purchase. Understanding these factors can assist businesses in
developing targeted marketing strategies, optimizing promotional activities, and improving
customer satisfaction to enhance the overall sales performance and profitability.

2.3. Metrics

2.3.1. C-Index

The C-index (also known as the concordance index or the area under the receiver operating
characteristic curve) is a widely used metric in survival analysis and medical research to assess
the performance of predictive models that estimate the likelihood of an event occurring over a
given time period.

The C-index is generated using the rankings of anticipated event occurrence probability for each
participant in a dataset. It calculates the percentage of pairings of people in whom the person with
the higher anticipated probability experienced the event before the person with the lower
projected probability. In other words, it assesses a predictive model's capacity to rank people in
order of their likelihood of experiencing the event of interest.
The C-index scales from 0 to 1, with 0.5 representing random prediction and 1 indicating perfect
prediction. In medical research, a C-index value of 0.7 or above is considered satisfactory
performance for a prediction model. Here is the formula of non-censored data C-Index.

Σij 1 𝑇𝑗 <𝑇𝑖 . 1𝜂𝑗 >𝜂𝑖 . 𝛿𝑗

𝐶 − 𝑖𝑛𝑑𝑒𝑥 =
Σij 1 𝑇𝑗 <𝑇𝑖 . 𝛿𝑗

𝜂𝑖 , 𝑡ℎ𝑒 𝑟𝑖𝑠𝑘 𝑠𝑐𝑜𝑟𝑒 𝑜𝑓 𝑎 𝑢𝑛𝑖𝑡 𝑖

1 𝑇𝑗 <𝑇𝑖 = 0 𝑖𝑓 𝑇𝑗 < 𝑇𝑖 𝑒𝑙𝑠𝑒 0
1𝜂𝑗 <𝜂𝑖 = 0 𝑖𝑓 𝜂𝑗 < 𝜂𝑖 𝑒𝑙𝑠𝑒 0
𝛿𝑗 , 𝑟𝑒𝑝𝑟𝑒𝑠𝑒𝑛𝑡𝑠 𝑤ℎ𝑒𝑡ℎ𝑒𝑟 𝑡ℎ𝑒 𝑣𝑎𝑙𝑢𝑒 𝑖𝑠 𝑐𝑒𝑛𝑠𝑜𝑟𝑒𝑑 𝑜𝑟 𝑛𝑜𝑡

3. RESULTS
The Kaplan-Meier curve presented below depicts the survival probability of time to purchase for
a cohort of online buyers. In this context, the x-axis represents the duration of time, while the y-
axis signifies the probability of surviving or completing a purchase. At the beginning of the
observation period, all buyers are considered to have a survival probability of 1, indicating that
they are all active participants.

6
 International Journal of Data Mining & Knowledge Management Process (IJDKP), Vol.13, No.4/5, September 2023

As time progresses, some buyers may not complete their purchases within a specific time frame.
In the context of this analysis, we refer to these incomplete purchases as "deaths."1 Consequently,
the survival probability gradually decreases over time, reflecting the diminishing likelihood of
completing a purchase within the given timeframe.

From a microeconomic perspective, the decreasing survival probability for time to purchase
suggests various factors at play. These may include consumer preferences, market conditions,
competitive influences, or external events that impact the decision-making process of buyers.
Microeconomists would delve deeper into these factors to gain insights into the underlying
dynamics affecting the time to purchase.

Figure 1: Kaplan-Meier survival curve.

Figure 2: Kaplan-Meier survival curve considering online behavior and gender.

1
In survival analysis, the term "events" or "deaths" is commonly used to refer to the occurrence of the event
of interest, which in this case would be completing a purchase. The survival probability, which represents
the likelihood of surviving (or completing a purchase) beyond a certain time point, naturally decreases over
time as more events or incomplete purchases occur.
7
 International Journal of Data Mining & Knowledge Management Process (IJDKP), Vol.13, No.4/5, September 2023

Figure 3: Kaplan-Meier survival curve considering online behavior and interests.

The figure above illustrates the changes in the risk of time to purchase over a given period.
Initially, there were 395 instances of purchases at time 1.07, with one purchase occurring. This
implies a high survival probability of 0.99747, indicating that most purchases were successfully
completed. However, as time progressed to 4.07, the survival probability slightly decreased to
0.99494, indicating two additional instances of time to purchase.

From a microeconomic perspective, this decreasing survival probability suggests an increase in

the risk of time to purchase over time. As we reach time 42.37, there were 382 instances
remaining, but 13 purchases had occurred during the observation period. Consequently, the
survival probability declined further to 0.96456, signifying a higher risk of delayed purchases.

Beyond time 42.37, the survival probability continued to rapidly decrease, suggesting a
significant increase in the risk of time to purchase. This can be attributed to various factors such
as changes in consumer behavior, market conditions, or competitive forces. A microeconomic
analysis would delve deeper into these factors to understand the underlying causes of the
increasing risk.

Analyzing the figure above, we can observe how the survival probability decreases as time
progresses. For instance, at time 32, there were 664 instances at risk, with 18 events (time to
purchase). The survival probability stood at 0.974, indicating a relatively low risk during this
period. However, as time advanced to 350, only 23 instances remained, and 51 purchases
occurred, resulting in a survival probability of 0.183. This indicates a significantly higher risk of
delayed purchases at that point.

By examining these trends and patterns, marketers can gain valuable insights into the changing
risk of time to purchase. This understanding allows them to adapt their strategies, refine targeting
approaches, and optimize marketing efforts to mitigate the risk and improve conversion rates
over time.

8
 International Journal of Data Mining & Knowledge Management Process (IJDKP), Vol.13, No.4/5, September 2023

Figure 4: Relative weight of each variable based on MTLR model.

In the figure above, the variables that have the most weight to define the moment of a purchase
are:

PurchaseHistory: The weight associated with the customer's purchase history has relatively
larger values compared to other variables. It suggests that a customer's past purchase behavior is
a strong predictor of the time to buy. Positive weights indicate that a higher purchase history is
associated with a longer time to buy.

PromotionsDiscounts: The weight for promotions and discounts also has notable values. It
indicates that offering promotions and discounts can influence the time to buy. Positive weights
suggest that promotions and discounts are associated with a longer time to buy.

CustomerExperience: The weight for customer experience has a noticeable impact. It suggests
that positive customer experiences can affect the time to buy. Positive weights indicate that a
positive customer experience is associated with a longer time to buy.

Age: The age weight, although smaller in magnitude compared to the previous variables, still
plays a role in predicting the time to buy. Positive weights imply that higher age values are
associated with a longer time to buy.

Gender, Income, MaritalStatus, Location, OnlineBehavior, and Interests: These variables have
smaller weights compared to the principal variables mentioned above. While they still contribute
to the prediction, their impact might be relatively less significant. However, it is important to note
that the relative importance of these variables may vary in different contexts.

3.1. What about Online Behavior and Interests?

When considering the time to buy a product basket, the weights associated with
"OnlineBehavior" and "Interests" in the analysis output can provide insights into the factors that
influence the duration it takes for customers to make a purchase. Here's how these variables can
be interpreted in relation to the time to buy:

9
 International Journal of Data Mining & Knowledge Management Process (IJDKP), Vol.13, No.4/5, September 2023

OnlineBehavior: the weights associated with "OnlineBehavior" indicate the impact of various
online behaviors on the time to buy a product basket. Positive weights (e.g., 0.009428, 0.011134,
0.015881) suggest that certain online behaviors are associated with a longer time to buy. This
implies that customers who exhibit these specific online behaviors might spend more time
researching, comparing options, or engaging with digital channels before making a purchase
decision.

On the other hand, negative weights (e.g., -0.00424, -0.01534, -0.01564) indicate that certain
online behaviors are associated with a shorter time to buy. This suggests that customers
displaying these behaviors might have a faster decision-making process, possibly due to higher
confidence or familiarity with the product or brand. Analyzing specific online behaviors included
in this variable can provide deeper insights into customers' digital engagement patterns and their
impact on the time to purchase a product basket.

Interests: the weights associated with "Interests" offer insights into how customers' specific
interests or preferences influence the time it takes for them to buy a product basket. Positive
weights (e.g., 0.00294, 0.00475, 0.01321) indicate that certain interests are associated with a
longer time to buy. This suggests that customers who have these specific interests might require
more time to make a purchase decision, potentially due to the need for extensive research or
consideration. Negative weights (e.g., -0.0131, -0.00931) suggest that certain interests are
associated with a shorter time to buy.

This implies that customers with these interests might have a quicker decision-making process,
possibly due to strong preferences or immediate needs related to the product basket. Analyzing
the specific interests captured in this variable can help identify customer segments and tailor
marketing strategies to align with their preferences, potentially influencing their time to buy.

By considering both OnlineBehavior and Interests, marketers can gain a more comprehensive
understanding of the factors that contribute to the time it takes for customers to purchase a
product basket. Leveraging these insights, marketers can develop targeted marketing campaigns,
personalized recommendations, and customer experiences that align with customers' online
behaviors and interests, ultimately optimizing the timing and relevance of their efforts to drive
conversions.

3.2. Model Comparison

The paper analyzed the performance of different machine learning survival models in predicting
time-to-purchase using a set of relevant variables. The concordance index (C-index) was used to
compare the predictive power of different models.

Based on the findings presented in the study, the performance of various survival models in
predicting time to purchase was assessed using the C-index metric. Among the models
considered, the DeepSurv model demonstrated the highest C-index value of 0.888524, indicating
superior predictive accuracy compared to the other models. The RandForest model achieved the
second-highest C-index of 0.724273, followed by the MTLR model with a C-index of 0.531151.
The Cox model yielded a C-index of 0.507152, while the KarnelSVM model performed least
effectively, obtaining the lowest C-index of 0.571.

These results suggest that the DeepSurv model outperformed the other models in predicting time
to purchase. The RandForest and MTLR models also exhibited reasonably good predictive
performance, although not as high as the DeepSurv model. This information holds significance

10
 International Journal of Data Mining & Knowledge Management Process (IJDKP), Vol.13, No.4/5, September 2023

for microeconomists, enabling them to understand and forecast the likelihood of purchase
completion, aiding in decision-making processes and risk mitigation strategies.

Figure 5: results from different machine learning models.

The variables examined in this study, such as Gender, Income, Location, Purchase History,
Online Behavior, Interests, Promotions Discounts and Customer Experience, offer valuable
insights into the factors contributing to understand what the moment of purchase will be, and its
associated probability.

By comprehending the impact of these variables, stakeholders in the field of survival analysis can
take appropriate measures to reduce the likelihood of incomplete purchases, leading to more
accurate predictions and informed decision-making. As such, this research contributes to
advancing the field of survival analysis in the domain of marketing and microeconomics.

4. CONCLUSION
The variables analyzed in this study provide valuable insights into the factors that influence the
time to purchase. Purchase history emerges as the most significant variable, indicating that a
customer's past purchase behavior strongly predicts the time it takes to make a purchase.
Promotions and discounts, along with positive customer experiences, also play notable roles in
influencing the duration. Age, although less impactful than the primary variables, still contributes
to the prediction.

Regarding the variables "OnlineBehavior" and "Interests," they offer further understanding of
customer behavior and preferences in relation to the time to buy. Certain online behaviors are
associated with longer decision-making processes, while others are linked to quicker purchase
decisions. Similarly, specific interests or preferences can either extend or shorten the time it takes
to make a purchase.

By considering both online behavior and interests, marketers can gain comprehensive insights
into customer segments and tailor their strategies accordingly. This knowledge can be leveraged
to develop targeted marketing campaigns, personalized recommendations, and customer

11
 International Journal of Data Mining & Knowledge Management Process (IJDKP), Vol.13, No.4/5, September 2023

experiences aligned with customers' online behaviors and interests, thereby optimizing
conversion rates.

Furthermore, the study demonstrates the effectiveness of machine learning survival models in
predicting the time to purchase. The DeepSurv model exhibited superior predictive accuracy,
followed by the RandForest and MTLR models. These models provide valuable tools for
microeconomists to forecast purchase completion and make informed decisions to mitigate risks.
Based on our results, the research proposes a number of possible next actions and areas for
development in order to increase our knowledge of the variables impacting the time-to-buy
decisions. Analyzing a bigger and more varied sample may give a wider viewpoint and boost the
results' generalizability. We may assess whether the observed associations remain true across
other settings by including data from numerous firms or sectors. While the present research
focused on a single organization, data from other companies or sectors would be useful for
comparing the outcomes. This would aid in identifying similar trends and determining if the
variables impacting purchase timing are constant across diverse situations.

Although the research looked at numerous major characteristics, there may be additional factors
that influence the timing to buy. Additional factors might include macroeconomic situation,
geographic region, or client preferences for certain product qualities. Extending the number of
variables examined may offer a more complete picture of the dynamics at work. Other customer-
related criteria, such as customer happiness, loyalty, or engagement, may give further insights
into their influence on the time to buy in addition to the variables examined. Incorporating these
variables may aid in determining whether they interact with or mediate the correlations reported
in the present investigation.

Validating the results using independent external datasets would be good. By comparing the
outputs across multiple datasets and establishing the stability of the correlations, this may assist
assure the robustness and dependability of the results. Combining quantitative and qualitative
research approaches, like as interviews or focus groups, may give a more in-depth insight of
consumers' underlying motives and decision-making processes. This qualitative data may help to
expand the understanding of quantitative findings and provide context for the results.

It is vital to emphasize the following restrictions that we examine at this time. The research was
done on a single firm, limiting the results' generalizability to other companies or sectors. It is
critical to recognize this restriction and evaluate the necessity for replication studies in various
situations. It is critical that the data utilized in the analysis be accurate and full. Assessing the
data's dependability and dealing with any possible data quality concerns, such as missing or
skewed data, should be a top focus.

The study intended to discover links between attributes and the moment of purchasing. Creating
causal links, on the other hand, requires more research, such as experimental designs or quasi-
experimental approaches. The study may not have considered all external factors that may
influence when to purchase, such as macroeconomic conditions, industry developments, or
competitors' strategies. Taking these external aspects into consideration in future studies may
provide a fuller understanding of the dynamics involved.

Researchers may continue to progress the discipline of survival analysis and increase our
knowledge of the variables influencing the time to buy by addressing these next steps and
constraints, allowing companies to make more informed choices and improve their marketing
tactics.

12
 International Journal of Data Mining & Knowledge Management Process (IJDKP), Vol.13, No.4/5, September 2023

REFERENCES
[1] Andersen, P. K., & Gill, R. D. Cox’s regression model for counting processes: A large sample study.
The Annals of Statistics,1982 10, 1100–1120.
[2] Bahari, T.F.; Elayidom, M.S. An Efficient CRM-Data Mining Framework for the Prediction of
Customer Behaviour. Procedia Computer Science 2015, 46, 725–731.
[3] Balducci, B.; Marinova, D. Unstructured data in marketing. Journal of the Academy of Marketing
Science 2018, 46, 557–590. Publisher: Journal of the Academy of Marketing Science
[4] Briesch, R.; Rajagopal, P. Neural network applications in consumer behavior. Journal of Consumer
Psychology 2010, 20, 381–389. Publisher: Society for Consumer Psychology,
[5] Carling, T., Pan, D., Ariyan, S., Narayan, D., & Truini, C. Diagnosis and treatment of interval
sentinel lymph nodes in patients with cutaneous melanoma. Plastic and Reconstructive Surgery,
2077 119(3), 907–913.
[6] Çelik, A Research on Machine Learning Methods and Its Applications. Journal of Educational
Technology and Online Learning 2018.
[7] Chang, S.-H., & Wang, M.-C. Conditional regression analysis for recurrence time data. Journal of
the American Statistical Association, 1999 94(448), 1221–1230.
[8] Chava, S., & Jarrow, R. A. Bankruptcy prediction with industry effects. Review of Finance, 2004
8(4), 537–569.
[9] Chen, S.H. The gamma CUSUM chart method for online customer churn prediction. Electronic
Commerce Research and Applications 2016, 17, 99–111. Publisher: Elsevier B.V.,
[10] Duarte, V., Zuniga-Jara, S., & Contreras, S. Machine Learning and Marketing: A Literature Review.
2018. Available at SSRN 4006436.
[11] Dzyabura, D.; Hauser, J.R. Active Machine Learning for Consideration Heuristics Active Machine
Learning for Consideration
Heuristics 2011.
[12] Erevelles, S.; Fukawa, N.; Swayne, L. Big Data consumer analytics and the transformation of
marketing. Journal of Business Research 2016, 69, 897–904. doi:10.1016/j.jbusres.2015.07.001.
[13] Hauser, J.R. Consideration-set heuristics . Journal of Business Research 2014, 67, 1688–1699.
Publisher: Elsevier Inc.,
[14] Jessen, H.C.; Paliouras, G. Data Mining in Economics, Finance, and Marketing. In Machine
Learning and Its Applications; Goos, G.; Hartmanis, J.; van Leeuwen, J.; Paliouras, G.; Karkaletsis,
V.; Spyropoulos, C.D., Eds.; Springer Berlin Heidelberg: Berlin, Heidelberg, 2001; Vol. 2049, pp.
295–299. Series Title: Lecture Notes in Computer Science, doi:10.1007/3-540-44673-7_18.
[15] Koutanaei, F.N.; Sajedi, H.; Khanbabaei, M. A hybrid data mining model of feature selection
algorithms and ensemble learning classifiers for credit scoring. Journal of Retailing and Consumer
Services 2015, 27, 11–23. Publisher: Elsevier
[16] Lane, W. R., Looney, S. W., & Wansley, J. W. An application of the cox proportional hazards model
to bank failure. Journal of Banking & Finance, 1986 10(4), 511–531.
[17] Lee, H.; Shin, H.; seob Hwang, S.; Cho, S.; MacLachlan, D. Semi-Supervised Response Modeling.
Journal of Interactive Marketing 2010, 24, 42–54. Publisher: Direct Marketing Educational
Foundation, Inc.
[18] Lee, S.; Kim, W. Sentiment labeling for extending initial labeled data to improve semi-supervised
sentiment classification. Electronic Commerce Research and Applications 2017, 26, 35–49.
Publisher: Elsevier B.V.
[19] Leonardis, D. D., & Rocci, R. Assessing the default risk by means of a discrete – time survival
analysis approach. Applied Stochastic Models in Business and Industry, 2008,24(4), 291–306.
[20] Leong, R. W., Nguyen, N., Meredith, C. G., et al. In vivo confocal endomicroscopy in the diagnosis
and evaluation of celiac disease. Gastroenterology, 2008, 135(6), 1870–1876.
[21] Liu, X.; Singh, P.V.; Srinivasan, K. A structured analysis of unstructured big data by leveraging
cloud computing. Marketing Science 2016, 35, 363–388.
[22] Luoma, M., & Laitinen, E. K. Survival analysis as a tool for company failure prediction. Omega,
1991, 19(6), 673–678.
[23] Ma, L.; Sun, B. Machine learning and AI in marketing – Connecting computing power to human
insights. International Journal of Research in Marketing 2020, 37, 481–504.

13
 International Journal of Data Mining & Knowledge Management Process (IJDKP), Vol.13, No.4/5, September 2023

[24] Misra, K.; Schwartz, E.M.; Abernethy, J. Dynamic online pricing with incomplete information using
multiarmed bandit experiments. Marketing Science 2019, 38, 226–252.
doi:10.1287/mksc.2018.1129.
[25] Prentice, R. L., Williams, B. J., & Peterson, A. V. On the regression analysis of multivariate failure
time data. Biometrika, 1981, 68, 373–389.
[26] Schwartz, E.M.; Bradlow, E.; Fader, P. Customer Acquisition via Display Advertising Using Multi-
Armed Bandit Experiments. SSRN Electronic Journal 2013.
[27] Seng, J.L.; Chen, T. An analytic approach to select data mining for business decision. Expert
Systems with Applications 2010, 37, 8042–8057. doi:10.1016/j.eswa.2010.05.083.
[28] Sun, Z.; Guo, Q.; Yang, J.; Fang, H.; Guo, G.; Zhang, J.; Burke, R. Research commentary on
recommendations with side information: A survey and research directions. Electronic Commerce
Research and Applications 2019, 37, 100879. Publisher: Elsevier _eprint:
1909.12807, doi:10.1016/j.elerap.2019.100879.
[29] Tkáˇc, M.; Verner, R. Artificial neural networks in business: Two decades of research. Applied Soft
Computing 2016, 38, 788–804.doi:10.1016/j.asoc.2015.09.040.
[30] Twisk, J., Smidt, N., & de Vente, W. Applied analysis of recurrent events: a practical overview.
Journal of Epidemiology and Community Health,2005, 59, 706–710.
[31] Xia, K.; Gao, L.; Wang, L.; Li, W.; Li, X.; Ijomah, W. Service-oriented disassembly sequence
planning for electrical and electronic equipment waste. Electronic Commerce Research and
Applications 2016, 20, 59–68. Publisher: Elsevier B.V.
[32] Wei, L. J., Lin, D. Y., &Weissfeld, L. Regression analysis of multivariate incomplete failure time
data by modelling marginal distributions. Journal of the American Statistical Association, 1989,
84(408), 1065–1073.

AUTHOR
Diego Vallarino, PhD, I am an experienced C-level executive and Principal Data
Scientist with over 20 years of experience in the financial services and information
business industries. My expertise lies in business development using data science. I
have held prominent positions such as Global Head Data Discovery at Coface
Group, Chief Data Scientist at Scotiabank, Data Analytics Director at Equifax, and
Senior International Consultant for organizations like the World Bank, IADB, and
UNDP.

In 2023, I was awarded the Einstein Visa (EB1A Green Card) by the US government. I have been
recognized as a "Global Top 100 Data Visionary" by Truata in 2020 and received the Global One Equifax
Award in 2018. Currently, I also serve as a Board Member for UTEC University in Peru and am a full
member of CDO LatAm (Peru) since 2020 and LatinxInAI (San Francisco, US) since 2022. Additionally, I
am a member of The Economic History Society since 2021 and the South American Network of Applied
Economics since 2009.

As an author, I published the book "Innovation from the South" in 2005, and I have delivered over 40
international conferences worldwide.

My academic background includes a PhD in Cliometrics from Universidad Torcuato Di Tella (Argentina)
completed between 2016 and 2018. I also hold an MSc in (bio)Statistics from the University of Barcelona
(Spain) earned between 2020 and 2022, and an MSc in Data Management & Big Data from the same
university in 2015. I furthered my education with an MBA Full Time from Universidad Adolfo Ibáñez
(Chile) in 2002-2003 and obtained a BSc in Business & CPA from FCEA - UdelaR (Uruguay) between
1994 and 2000.

Throughout my career, I have pursued continuous learning through post-academic degrees and executive
education courses. Notable among them are SEE from Babson College (USA, 2007), GCP from The
Wharton School (USA, 2003), MOC in Microeconomics of Competitiveness from Harvard Business
School (USA, 2010), and Data Science Ethics from the University of Michigan (USA, 2018).

14