GRADUATE THESIS APPLICATION OF DEEP LEARNING MODELS IN FORECASTING SALES TIME SERIES DATA MAJOR: MANAGEMENT INFORMATION SYSTEM CODE: 7340405 STUDENT NAME: LAM THI THAO NGUYEN HO CH
INTRODUCTION TO THE RESEARCH TOPIC
Statement of the research problem
Amid a volatile global economy and intensified market competition, accurate sales forecasting enables retail businesses to predict future sales through data-driven analytics, serving as a pivotal decision-making tool across functions such as inventory management (including avoiding overproduction and reducing overstock), marketing, and workforce planning (Islam & Amin, 2020; H D Nguyen et al., 2021) By providing more precise cost projections, forecasting can boost profitability and overall performance (Huang et al., 2019; Islam et al., 2021; Sohrabpour et al., 2021; Verstraete et al., 2020) Inaccurate forecasts, however, can trigger significant losses through overstocking, missed sales opportunities, and wasted resources, harming the bottom line (Magrini, 2023).
Sales forecasting is the practice of estimating future sales by analyzing historical data, market trends, and other relevant factors to predict the quantity of products or services a company is likely to sell over a defined period, such as weeks, months, or years It blends quantitative and qualitative methods to generate insights that help businesses prepare for future demand (Isreal & Alonge, n.d.) Analyzing sales patterns—whether for total store sales or individual products—helps identify the major drivers of future sales Forecasting difficulty varies across product types: staples like milk tend to have stable consumption and are easier to forecast (Choprar & Meindl, 2001; Padilla et al., 2021), whereas fashion items or furniture are shaped by trends and seasonality, adding complexity to predictions.
Time series analysis is among the most common techniques for sales forecasting, especially in markets with clear seasonal or cyclical demand By examining time-ordered data points such as monthly or quarterly sales, businesses can uncover trends, seasonality, and long-term demand shifts, enabling more accurate forecasts and smarter inventory and revenue planning (MAGRINI, 2023).
Traditional forecasting models such as the Autoregressive Integrated Moving Average (ARIMA) have been widely adopted for time series forecasting because they can capture recurring patterns and short-term dependencies (Box et al., 2015) Exponential smoothing methods, including the Holt-Winters seasonal model, are effective for seasonally adjusted forecasts and perform well in stable, predictable environments (Hyndman & Athanasopoulos, 2018) However, these approaches rely on strong assumptions of linearity and stationarity, which makes them less effective when time series exhibit nonlinear patterns, noise, or long-range dependencies that are common in many modern datasets.
In recent years, deep learning (DL)—a powerful subset of machine learning (ML)—has surpassed traditional ML methods across many tasks, driven by rapid methodological breakthroughs These advances have extended to time series forecasting, where DL can model complex, high-dimensional temporal patterns that are often beyond the reach of traditional approaches Core DL architectures such as Long Short-Term Memory (LSTM), Gated Recurrent Units (GRU), and Temporal Convolutional Networks (TCN) excel at learning intricate temporal dependencies with little manual feature engineering and without rigid assumptions about the data structure.
Although interest in applying deep learning to sales time series forecasting is rising, detailed investigations in this area remain relatively scarce, particularly for emerging markets and real-world business data Consequently, there are few empirical studies focused on evaluating and comparing these models for sales forecasting tasks, leaving gaps in understanding their relative strengths and limitations across diverse market conditions.
Research objectives
The primary objective of this study is to investigate the application of deep learning models in sales time series forecasting and evaluate their effectiveness compared to traditional statistical methods
- Select and implement deep learning models such as Long Short-Term Memory (LSTM), Gated Recurrent Units (GRU), and Temporal Convolutional Networks (TCN) for sales time series forecasting
- Apply deep learning models to actual sales datasets, specifically those obtained from publicly available sources such as Kaggle
- Compare the forecasting performance of deep learning models with traditional statistical methods
- Compare the strengths and limitations of traditional and deep learning models in handling real-world sales data challenges
- Provide practical recommendations for businesses in optimizing sales forecasting and inventory management based on the study findings.
Research questions
- Which deep learning models are appropriate for forecasting sales time series data, and how can they be effectively implemented?
- How do LSTM, GRU, and TCN models perform when applied to actual sales data?
- How does the forecasting performance of deep learning models compare with that of traditional statistical methods?
- What are the strengths and limitations of traditional and deep learning models in handling real-world sales data challenges?
- What practical recommendations can be provided for businesses in optimizing sales forecasting and inventory management based on the study findings?
Subject and scope of the study
This study examines the use of deep learning architectures—LSTM, GRU, and Temporal Convolutional Networks (TCN)—for forecasting sales time series data It benchmarks these models against traditional statistical methods, including ARIMA, SARIMA, and Exponential Smoothing, to assess their relative predictive accuracy The aim is to determine whether deep learning approaches can deliver more accurate forecasts in retail environments, supporting better inventory planning, demand forecasting, and revenue optimization.
Subject of the research: Sales transaction data from a retail business, focusing on key product categories or items within the time series dataset
This dataset comprises about two years of sales time series data with daily granularity, capturing fields such as date, quantity, unit price, and product details Its core emphasis is daily aggregation to reveal trends, seasonality, and anomalies, while enabling forecasting for selected high-volume items to inform inventory planning and revenue optimization By focusing on item-level performance over time, the data supports accurate demand forecasting, price sensitivity analysis, and product segmentation The inclusion of product metadata alongside core metrics enhances analysis for merchandising, pricing strategy, and supply chain decision making.
Contribution of the study
This study contributes to both academic research and practical applications:
This study advances the theoretical understanding of time series forecasting by demonstrating that deep learning architectures can effectively capture complex, non-linear patterns in sales data Through empirical comparisons with traditional models, the results highlight how deep learning excels in managing seasonality, sparsity, and volatility These findings enrich the forecasting literature and offer insights for future research on model scalability and interpretability in data-driven decision-making.
This study provides actionable insights for businesses aiming to improve sales forecasting accuracy through advanced deep learning techniques, offering comparative evaluations of multiple models so organizations can select approaches aligned with their data characteristics and operational needs Implementing these insights can enhance inventory management, optimize resource allocation, and strengthen decision-making, especially for retailers and other firms operating in dynamic markets with access to time-series data.
Structure of the study
Chapter 1: Introduction to the research topic
1.1 Statement of the research problem
1.4 Subject and scope of the study
2.2 Overview of time series forecasting in business
2.3 Neural network and deep learning
3.2 Data collection and dataset description
Chapter 4: Research results and discussion
4.1 Forecasting results and model evaluation
4.2 Comparison and discussion of model performance Chapter 5: Conclusion and recommendations
5.2 Practical implications and recommendations 5.3 Limitations and directions for future research
THEORETICAL BACKGROUND
Review of related literature
3.2 Data collection and dataset description
Chapter 4: Research results and discussion
4.1 Forecasting results and model evaluation
4.2 Comparison and discussion of model performance Chapter 5: Conclusion and recommendations
5.2 Practical implications and recommendations 5.3 Limitations and directions for future research
Chapter 2 presents the fundamental concepts of sales time series forecasting and the theoretical foundations underpinning this study In addition, it reviews previous research related to both traditional statistical methods and deep learning models, thereby providing the background for the proposed research
Forecasting is the systematic process of predicting future events or conditions by analyzing patterns and uncovering trends in historical and current data (IBM, 2024) This approach informs organizational decision-making and supports key activities such as production scheduling, workforce allocation, transportation planning, and long-term strategic development (Hyndman & Athanasopoulos, 2018).
Forecasting is a core capability for any organization, enabling proactive planning and informed decision‑making across a wide range of functions and sectors It spans multiple fields—from business and public administration to environmental sciences, healthcare, economics, and politics—supporting budgeting, operations, policy development, resource allocation, and risk management.
Forecasting problems are typically categorized based on the length of the time horizon Short-term forecasting involves predictions over a few days to several months and is commonly used for operational decisions Medium-term forecasting usually covers a timeframe of one to two years, supporting tactical decisions, and Long-term forecasting extends over several years and informs strategic planning activities Short- and medium-term forecasting plays a critical role in supporting a range of organizational functions such as operations management, budgeting, and the evaluation of new research and development initiatives In contrast, long-term forecasting contributes significantly to strategic decision-making and long-range planning The effectiveness of short- and medium-term forecasting is largely attributed to its reliance on recognizing and modeling historical data patterns, which typically display stability and gradual change over time Due to this consistent nature of past data, statistical methods are very useful for short and medium-term forecasting (Montgomery et al., 2015)
One of the most prevalent and impactful applications of forecasting in business is sales forecasting A sales forecast represents the sales revenue a business expects to generate It provides an estimate of the amount of products or services the company anticipates selling over a specific period, such as a quarter or a year (Salesforce, 2023)
Sales forecasting enables businesses to establish attainable sales targets, optimize resource allocation, and adjust strategic plans in response to changing market conditions, helping organizations align operations with realistic goals and stay agile as market dynamics evolve (Pao & Sullivan, 2017).
Sales forecasts serve as a foundational input for multiple organizational functions, guiding finance and accounting in estimating costs, profits, and capital needs (Mentzer & Bienstock, 1998) The sales division relies on accurate projections to plan and deploy its workforce effectively For production and procurement, long-term forecasts inform facility and equipment planning, while short-term forecasts are essential for shaping the production schedule Marketing teams use forecasted sales to shape campaigns and assess their impact on demand Logistics operations leverage both short- and long-term forecasts to structure infrastructure and coordinate delivery and inventory needs (Doganis et al., 2006).
As markets become increasingly global and competition intensifies, optimizing an organization’s operational efficiency becomes a premium priority In a landscape where resources must be spread broadly and consumers have abundant choices, every efficiency gain can impact performance—achieving the right balance between supply and demand eliminates missed sales from stockouts and reduces costs from overstocking Sales forecasting uses patterns drawn from historical data to predict future demand, enabling informed actions such as reallocating inventory, diverting stock, and adjusting production levels (Pao & Sullivan, 2017).
Although the forecasting process may vary between organizations, it generally follows these steps (IBM, 2024):
Defining the forecasting objective is the first step in any predictive effort Begin by identifying the specific business metric you want to predict—such as sales revenue, product demand, or customer churn—and specify the time horizon and granularity for the forecast This stage also involves outlining the key assumptions behind the forecast and listing the relevant variables that will drive the model, including historical data, market trends, seasonality, promotions, and other economic indicators Clear goal setting, scope, and data inputs lay the foundation for accurate predictions and transparent evaluation of the forecast's performance.
Data collection and preparation involve gathering historical data and evaluating it for quality and relevance to support reliable outcomes This process may draw on internal company records, external market data, or third-party information, and selecting datasets that align with the analysis goals helps produce accurate results.
Choosing a forecasting method means selecting a forecasting approach that fits the specific business case and the performance metrics you want to optimize, while also aligning with the relevant variables, underlying assumptions, and the distinctive characteristics of your datasets This alignment ensures the method can produce reliable predictions that reflect data structure, seasonality, trends, and data quality, supporting better data‑driven decisions When evaluating options, compare multiple methods against the same metrics, assess how well each handles variability and noise, and verify that its assumptions match the data-generating process.
- Generating the forecast: The selected forecasting method is applied to analyze the data, upon which the forecast is subsequently constructed based on the results of this analysis
- Validating and optimizing the forecast: The accuracy and reliability of the forecast are assessed Adjustments or refinements may be applied to enhance predictive performance and minimize errors
Forecasts are presented in a clear, accessible format, often leveraging data visualization tools to illustrate key trends and metrics This presentation enables stakeholders to interpret outcomes quickly and make informed strategic or operational decisions.
Although forecasting is applied in various contexts, there are only two broad types of forecasting techniques: qualitative forecasting and quantitative forecasting (Montgomery et al., 2015)
Qualitative forecasting is generally applied in cases where historical data is either unavailable or irrelevant to the forecasting objective (Hyndman & Athanasopoulos,
Qualitative forecasting, unlike statistical models, relies on human expertise, subjective judgment, and insights from stakeholders such as consumers, industry specialists, and senior executives To ensure consistency, it often uses structured evaluation methods that systematically convert qualitative information into quantitative data (IBM, 2024).
One widely recognized qualitative forecasting technique is the Delphi Method (Montgomery et al., 2015), which entails multiple rounds of anonymous surveys answered by experts Responses are aggregated and refined in subsequent rounds until a consensus is reached, a process that minimizes bias and ensures balanced input from all contributors (IBM, 2024).
Market research is a common qualitative method in which firms collect primary data from consumers through surveys, focus groups, and interviews The feedback gathered informs sales projections, marketing strategies, and product development, while also converting subjective assessments into structured ratings or categorical data This enables decision-makers to make informed choices even when historical data is limited or unavailable (IBM, 2024).
RESEARCH METHODOLOGY
Evaluation metrics
Chapter 4: Research results and discussion
4.1 Forecasting results and model evaluation
4.2 Comparison and discussion of model performance Chapter 5: Conclusion and recommendations
5.2 Practical implications and recommendations 5.3 Limitations and directions for future research
Chapter 2 presents the fundamental concepts of sales time series forecasting and the theoretical foundations underpinning this study In addition, it reviews previous research related to both traditional statistical methods and deep learning models, thereby providing the background for the proposed research
Forecasting is the systematic process of predicting future events or conditions by analyzing patterns and uncovering trends in historical and current data This approach underpins organizational decision-making by supporting activities such as production scheduling, workforce allocation, transportation planning, and long-term strategic development.
Forecasting is a core activity in any organization, underpinning a broad spectrum of business functions and sectors Its reach extends across multiple fields, including business, public administration, environmental sciences, healthcare, economics, and politics, underscoring forecasting’s cross-disciplinary relevance and value.
Forecasting problems are typically categorized by horizon length: short-term forecasting covers predictions from a few days to several months and supports operational decisions; medium-term forecasting spans about one to two years and informs tactical decisions; long-term forecasting extends over several years and underpins strategic planning Short- and medium-term forecasts play a key role in operations management, budgeting, and evaluating new research and development initiatives, while long-term forecasts contribute to strategic decision-making and long-range planning Their effectiveness for short- and medium-term horizons stems largely from recognizing and modeling historical data patterns, which tend to be stable with gradual change over time, making statistical methods particularly useful for these horizons (Montgomery et al., 2015).
Sales forecasting is one of the most prevalent and impactful applications of forecasting in business A sales forecast projects the revenue a company expects to generate and estimates the quantity of products or services it anticipates selling over a defined period, such as a quarter or a year (Salesforce, 2023).
Sales forecasting serves as a tool for businesses to establish achievable sales targets, allocate resources efficiently, and refine strategic plans in response to shifting market conditions By anticipating demand and trends, it enables better budgeting, capacity planning, and decision-making, helping organizations stay agile as the market evolves (Pao & Sullivan, 2017).
Sales forecasts act as a foundational input across the organization In finance and accounting, they are used to estimate costs, profits, and capital requirements; the sales division relies on accurate projections to manage and deploy its workforce; for production and procurement, long-term forecasts inform facility and equipment planning, while short-term forecasts are crucial for arranging the production schedule Marketing uses forecasted sales to shape campaigns and evaluate their impact on demand, and logistics benefits from both short- and long-term forecasts to structure infrastructure, coordinate deliveries, and manage inventory (Mentzer & Bienstock, 1998; Doganis et al., 2006).
Global markets are increasingly competitive, making operational efficiency a strategic priority for any organization As firms deploy resources across broader geographies and consumers face a flood of choices, even marginal efficiency gains can create meaningful competitive differentiation Matching product supply to demand minimizes stockouts and curbs the costs of overstocking, while sales forecasting uses patterns from historical data to predict future demand and guide actions such as reallocating inventory, redirecting supply, and adjusting future production levels (Pao & Sullivan, 2017).
Although the forecasting process may vary between organizations, it generally follows these steps (IBM, 2024):
Defining the forecasting objective begins with identifying the precise business metric to predict—such as sales revenue, product demand, or customer churn—and articulating the desired outcome This step also sets out the key assumptions behind the forecast and specifies the relevant variables that will inform the model By clarifying the objective, data requirements, and potential constraints, you create a solid foundation for model selection, accuracy evaluation, and actionable insights that align with strategic goals.
Data collection and preparation start with gathering historical data and evaluating its quality and relevance Sources include internal company records, external market data, and third-party information, with each source checked for accuracy, completeness, and timeliness Selecting the right datasets is a key driver of reliable outcomes, because data quality and source choice directly influence analytics credibility and the usefulness of insights.
Selecting the forecasting method means choosing an approach that aligns with the business objective and the chosen performance metric while also fitting the relevant variables, assumptions, and data characteristics of the datasets The right method reflects the business context, leverages the data structure, and accommodates factors such as data quality, forecast horizon, and variable relationships By evaluating these elements—business case, performance metrics, variables, assumptions, and dataset attributes—you can pick a forecasting technique that delivers accurate, actionable insights.
- Generating the forecast: The selected forecasting method is applied to analyze the data, upon which the forecast is subsequently constructed based on the results of this analysis
- Validating and optimizing the forecast: The accuracy and reliability of the forecast are assessed Adjustments or refinements may be applied to enhance predictive performance and minimize errors
The forecast is presented in a clear, accessible format, often supported by data visualization tools, allowing stakeholders to interpret the outcomes quickly This clarity enables informed strategic and operational decision-making based on solid insights.
Although forecasting is applied in various contexts, there are only two broad types of forecasting techniques: qualitative forecasting and quantitative forecasting (Montgomery et al., 2015)
Qualitative forecasting is generally applied in cases where historical data is either unavailable or irrelevant to the forecasting objective (Hyndman & Athanasopoulos,
Rather than relying on statistical models, qualitative forecasting hinges on human expertise, subjective judgment, and insights from consumers, industry specialists, or senior executives To ensure consistency, it uses structured evaluation methods that systematically translate qualitative information into quantitative data (IBM, 2024).
Among qualitative forecasting methods, the Delphi Method is a widely used approach It relies on several rounds of anonymous questionnaires completed by subject-matter experts, with each round’s responses aggregated and fed back for refinement Through this iterative process, expert opinions converge toward a consensus, helping to reduce bias and ensure balanced input from all contributors (Montgomery et al., 2015; IBM, 2024).
Market research is a key qualitative method in which firms collect primary data from consumers through surveys, focus groups, or interviews The feedback gathered informs sales projections, shapes marketing strategies, and guides product development Qualitative methods convert subjective assessments into structured ratings or categorical data, helping decision-makers make informed choices even when historical data are scarce (IBM, 2024).