Forecasting is critically important in the aviation industry for analyzing booking trends, demand forecasting, revenue maximization, and strategic planning. Inaccurate forecasts lead to capacity losses, increased operating costs, and financial drainage for airlines.
This study employs multiple time series forecasting methods including ARIMA, SARIMA, ETS, X11+ETS, X11+ARIMA, STL+ARIMA, STL+RW-drift, and the machine learning technique LSTM on monthly passenger data of Biman Bangladesh Airlines from January 2017 to July 2025. The time series exhibits non-stationarity, upward trend, moderate seasonal strength (Fs = 0.4115), and structural breaks due to COVID-19.
STL decomposition with ARIMA and Random Walk with Drift methods performed optimally for traffic movement forecasting. Seasonal patterns significantly influence revenue and passenger volumes, requiring capacity adjustments. LSTM captures long-term dependencies but underperforms with smaller datasets. Approximately 41% of data variation is attributable to seasonal components. STL+ARIMA achieved the best performance with RMSE of 13,016.25 and MAPE of 4.68% on test data. Seasonal sub-series analysis revealed strong seasonal patterns in air passenger traffic and it depends on direction of the movement of the journey and month of the year.
The study demonstrates that decomposition methods combined with traditional forecasting techniques provide superior accuracy for airline passenger forecasting. The findings offer practical guidance for model selection based on data complexity, computational resources, and application requirements, enabling smarter strategic planning through predictive, data-backed intelligence. Accurate forecasting about market is required to optimize use of capacity of the equipment and render better services to the customers.
Accurate forecasting is not merely a technical exercise but a strategic imperative in the airline industry. The thin profit margins, perishable inventory, and intense competition characteristic of the industry make forecasting accuracy a critical determinant of financial performance.
Table of Contents
1.0 Introduction
1.1 Background of the study
1.2 Rationale of the Study
1.3 Project Scoping and Data Acquisition
1.4 Problem Statement and Motivation
1.5 Research Questions
1.6 Significance of the Study
1.7 Aviation Economy of Bangladesh
1.8 Organization of the Thesis
1.9 Research Objectives
2.0 Literature Review
2.1 Theoretical Background
2.2 Time Series Modeling Approaches
2.2.1 Traditional Statistical Methods
2.2.1.1 Transformation
2.2.1.2 Box Cox Transformation
2.2.1.3 Seasonal Sub-sirs plot
2.2.1.4 Residual Check
2.2.1.5 ARIMA
2.2.2 Exponential Smoothing Method
2.2.3 ETS Model
2.3 Advance Forecasting Approaches
2.3.1 Decomposition Methods
2.4 Forecasting Methods and Model
2.4.1 Machine Learning Model
2.4.2 Neural Network in Time Series
2.5 Related Empirical studies
2.5.1 Global context
2.5.2 Regional Context
2.6 Comparative Studies
2.7 Statistical Methods Applied in Similar Context
2.8 Research Gap
3.0 Research Methodology
3.1 Research Design
3.2.1 Data Collection
3.2.1 Data Sources
3.2.2 Data Characteristics
3.3 Exploratory Data Analysis (EDA)
3.3.1 Descriptive statistics
3.3.2 Time Series Plot
3.3.3 Correlation and Heat Map
3.3.4 Stationarity check
3.3.5 Seasonality Strength
3.3.6 Seasonal sub-series plot
3.4 Model Development
3.4.1 Traditional Model
3.4.1.1 ARIMA and SARIMA Model
3.4.2 Decomposition based Time series
3.4.3 Machine Learning Approaches
3.5 Model Evaluation Framework
3.6 ETS Model Development for Forecasting
3.7 Decomposition and Forecasting Model Development
3.7.1 X11 + ARIMA and X11+ETS
3.7.2 STL +NAIVE
3.7.3 STL + Random Walk with Drift
3.7.4 STL + ARIMA
3.8 Long short-term memory (LSTM)
3.8.1 Plot time series in LSTM
3.8.2 Prepare data for LSTM
3.8.3 Model build up
3.8.4 Train the model
3.9.5 Forecasting by LSTM
4.0 Result and Analysis
4.1 Exploration Data Analysis Findings
4.1.1 Data characteristics and distribution
4.1.2 Trend Analysis
4.1.3 Seasonal pattern identification
4.1.4 Stationarity and Correlation Analysis
4.2 Model Performance Comparison
4.2.1 Traditional Model Performance
4.2.2 Decomposition based Approaches
4.2.3 Advanced Machine Learning Performance
4.1 Model Evaluation and Cross validation
4.1.1 Comparison AIC
4.1.2 Accuracy
4.1.3 Determine the best-performing model
4.1.4 Comparison forecasting in Time series
5.0 Forecasting, Interpretation, and Reporting
5.1.1 Final Model
5.2 Prediction Intervals
5.3 Business Insights
5.4 Recommendations
6.0 Conclusion
Objectives & Topics
The primary objective of this study is to identify and evaluate the most accurate time series forecasting model for predicting monthly passenger volume and sales revenue for Biman Bangladesh Airlines. Operating in a highly volatile and seasonal industry with narrow profit margins and perishable inventory, the airline requires rigorous demand modeling to support flight scheduling, fleet allocation, and revenue management. The research specifically investigates whether decomposition-based statistical methods outperform classical univariate models and recurrent neural network architectures, such as Long Short-Term Memory (LSTM) networks, particularly under constraints of limited historical data and structural breaks caused by the COVID-19 pandemic.
- Comparative evaluation of forecasting methodologies: Classical models (ARIMA, SARIMA, ETS), decomposition hybrids (X11, STL combined with ARIMA and ETS), and deep learning (LSTM).
- Empirical examination of the Biman Bangladesh Airlines data ecosystem from January 2017 to July 2025.
- Analysis of seasonal variation, cyclical peaks, and structural breaks induced by external shocks such as COVID-19.
- Evaluation of aviation commercial metrics, including Revenue Passenger Kilometers (RPK), Blended Fare, and Blended Yield.
- Strategic recommendations for capacity management, fleet planning, dynamic pricing, and operational decision-making.
Excerpt from the Book
1.1 Background of the Study
Biman Bangladesh Airlines, the national flag carrier of Bangladesh, has demonstrated significant financial performance with total revenue of 10,575 crore taka in FY 2023-24, serving 3.3 million passengers. The subsequent fiscal year 2024-25 shows further growth with revenue reaching Tk 11,631.37 crore, 3.4 million passengers, and 43,918 tons of cargo, achieving a cabin factor of 82%. To sustain this growth trajectory and enhance market share, accurate forecasting of traffic movement is imperative.
The air transport sector significantly contributes to Bangladesh's economy. According to IATA, the aviation sector employed 29,200 people directly and 474,600 indirectly in 2023, contributing a total of USD 5.3 billion to GDP and generating USD 2.4 billion in direct economic output.
Aviation demand is influenced by multiple factors including economic conditions, geographical considerations, service quality, airfare dynamics, and market factors. The industry exhibits complex patterns of seasonality, time-of-day variations, equipment type preferences, and customer behavior. With advancements in machine learning techniques and advanced analytics, there exists a significant opportunity to leverage data science for developing more responsive, accurate, and profitable forecasting models.
Relationship is not always linear, so require to study of non-linearity, that can be studied by Activation functions of neural networks. In this study trying to choose of appropriate model in comparison of SARIMA, ETS and LSTM of RNN machine learning technique. Accurate forecasting is required for flight scheduling, fleet planning, capacity measurement, pricing, route expansion, SPA and Code share agreement and sales revenue maximization etc.
This study focuses on primary target variables including Passenger Count (monthly total) and Sales Revenue (total, in BDT/USD). While potential predictor variables such as operating costs, fuel prices, GDP, holiday indicators, and exchange rates demonstrate high correlation with passenger numbers, this project primarily analyzes univariate time series to understand patterns like trends, seasonality, and noise within the single variable to make forecasts.
Chapter Overview
1.0 Introduction: Introduces the research context, industry significance, research questions, and the macroeconomic role of civil aviation in Bangladesh.
2.0 Literature Review: Provides the theoretical foundations of airline revenue management, univariate time series modeling, exponential smoothing, time series decomposition, and neural networks.
3.0 Research Methodology: Details the empirical data collection pipeline, exploratory data analysis framework, stationarity diagnostics, and model implementation parameters for statistical and deep learning architectures.
4.0 Result and Analysis: Analyzes the empirical results, presenting error metric comparisons across ARIMA, ETS, STL decomposition combinations, and LSTM networks.
5.0 Forecasting, Interpretation, and Reporting: Discusses the 12-month forward forecasts, prediction intervals, operational planning implications, and strategic decision support for airline management.
6.0 Conclusion: Summarizes the key empirical conclusions regarding decomposition superiority, data limitations in deep learning, and practical implications for airline operators.
Keywords
Time Series Forecasting, Airline Revenue Management, STL Decomposition, SARIMA, LSTM, Bangladesh Aviation, Seasonal Pattern, Yield, Available Seat Kilometer, Revenue Passenger Kilo, Blended Revenue, Blended Fare, IATA
Frequently Asked Questions
What is the core focus of this research study?
The study provides an empirical comparison of statistical time series models and machine learning methods to forecast monthly passenger counts and revenue for Biman Bangladesh Airlines.
What are the primary thematic areas analyzed in the document?
The thesis explores airline revenue management, time series decomposition (STL and X11), traditional autoregressive methods, deep learning via LSTM recurrent neural networks, and aviation market dynamics in Bangladesh.
What is the primary objective and research question addressed by the author?
The primary objective is to determine the best-fit forecasting model among classical, decomposition-based, and deep learning approaches to improve passenger demand prediction and revenue optimization under seasonal volatility.
What scientific methodologies and quantitative techniques are employed?
The author uses empirical time series methodologies including ARIMA, SARIMA with Box-Cox transformation, ETS state-space models, X11 and STL decomposition combined with ARIMA and Random Walk with Drift, as well as an LSTM neural network built with TensorFlow/Keras.
What key insights are presented in the main analysis section?
The main analysis demonstrates that seasonal decomposition combined with ARIMA (STL+ARIMA) achieves superior forecasting accuracy (RMSE of 13,016.25 and MAPE of 4.68%), outperforming pure statistical models and deep learning architectures on small, highly seasonal datasets.
Which keywords best characterize the scope of the publication?
The publication is characterized by keywords including Time Series Forecasting, Airline Revenue Management, STL Decomposition, SARIMA, LSTM, Bangladesh Aviation, Blended Yield, and RPK.
Why did the LSTM neural network underperform relative to the STL+ARIMA model?
While the LSTM network demonstrated strong point forecasting capabilities on training samples, its generalization was constrained by the limited size of the monthly dataset (103 observations), where traditional decomposition-based statistical methods excel.
How does seasonality specifically impact Biman Bangladesh Airlines' passenger traffic?
Seasonal analysis revealed that seasonal components account for approximately 41% of data variation, characterized by predictable surges during December–January and July–August (driven by winter travel and Eid holidays) followed by pronounced lean periods in May and June.
What counterintuitive relationship was discovered between passenger numbers and airline yield?
The correlation analysis identified a moderate negative correlation (-0.236) between total passenger volume and blended yield, indicating that passenger increases do not automatically guarantee proportional revenue expansion, largely due to a low proportion of premium business cabin travelers.
- Quote paper
- Promitosh Talukder (Author), 2025, Forecasting Passenger and Sales Revenue for an Airline Using a Comparative Analysis of Different Time Series Models, Munich, GRIN Verlag, https://www.grin.com/document/1746485
