Theses supervised by Prof. Dr. Ahmet Refik Güllü
11 theses · Boğaziçi University
Application of a multi period multi retailer price decision and inventory allocation model
Effective dynamic pricing strategies and optimal inventory allocation have emerged as crucial drivers for operational success for vendor-retailer systems under supply chain constraints and economic uncertainties. Joint problems optimizing profit have been extensively studied in the literature while a gap remains for applications within a unified model with retailer specific warehouse to retailer lead times, lost sales costs as a cost component and inventory capacity levels. In this thesis, an existing multi period multi retailer profit optimization model is extended with these considerations. Real FMCG company data are calibrated to perform an industry application on the extended model leveraging their distributor sales channel's relevant practices and operational advantages for dynamic strategies. Empirical analysis using company data reveals that retailer capacity constraints act as a limiting factor under increasing transportation costs, with profit impacts ranging up to 1.01%, depending on supply chain cost ratios. The impact of fully dynamic pricing strategy on total profit is found to be up to 1.61%. Experimental results using synthetic data indicate that increased unit lost sales costs and longer lead times correlate to broader range of prices and elevated prices during initial frozen periods. Total profit is seen to be decreased under higher unit lost sales costs. Results indicate that straightforward comparison of total profit among different lead times is not straightforward due to complex trade-offs between model parameters. Nevertheless, insufficient initial retailer inventories result in lower total profit for longer lead times as longer lead times make it difficult for the model to compensate lost sales. Longer lead times require earlier allocation decisions, potentially impacting profitability in dynamic business environments. Model extensions significantly influence pricing and allocation decisions during frozen periods, suggesting that flexible systems can effectively compensate total profit through short-term adjustments rather than wholesale strategy revisions.
Option pricing under stochastic interest rate
This study proposes a newly generated model to price options under a stochastic interest rate environment. The developed model introduces a numerical procedure with the arbitrage-free condition by working on a binomial tree. The model handles price changes in stocks and interest rates. Principally, it is assumed that the movements in the stock prices are defined by the Cox, Ross, Rubinstein (CRR) model. The CRR model proposes a numerical method to price options by assuming the interest rate is constant throughout the option's life. Moreover, the thesis claims that interest rates vary based on the Black, Derman, Toy (BDT) model. The BDT model defines the evolution of interest rates in the future. It presents a numerical procedure by using the binomial tree. Crucially, the interest rate is log-normally distributed in the BDT model; hence, the short rate cannot take a negative value. Also, the BDT model assumes that it has a mean-reverting property which means that the interest rate shows a tendency to converge to the average of interest rates in the long term. Additionally, this study utilizes the CRR and BDT model in order to derive a new option valuation framework. Also, the proposed model gives a numerical solution rather than an analytical formula due to the BDT model's structure. This thesis focuses on pricing the European options that can expire only on the maturity date. Furthermore, a group of options with different strike prices and different maturities is valued according to the developed model and the CRR model to observe interest rate impacts under two parameters: strike price and time-to-maturity. Finally, the estimated prices by both models are compared with actual-market prices to determine the accuracies of the models. Then, it is detected that the effect of the stochastic interest rate behavior on which maturities is significant.
Multiple queues with simultaneous arrivals
Queuing theory problems have been the topic of deep research owing to the fact that so many difficulties are in existence and their significance in real life cases can not be ignored. Those problems can be observed in numerous sectors such as telecommunications, airlines, logistics, hospitals, computing, production and inventory. Besides, speed is the key word in today's world because population is almost at the peak, thus demands or requests must be met as much as possible. However, our world has limited sources that is why there has to be some delays and queues. Additionally, game theory is one of the most important topics and it comes into prominence due to increasing competition in the world. There are lots of organizations which dwell in aforementioned sectors and they need to compete with each other to maximize their benefits. Just as in queuing theory, application of game theory spans the huge part of real life problems involving so much burden. So, there are abundance of works which dive into the distinct branches of game theory. In this study, both queueing theory and game theory are taken into consideration. We include the concept of game analysis, server rate optimization, multiple queues, loss systems and simultaneous arrivals at the same time whereas the studies in literature just focus on some of them. In our first case, we apply a game theoretic approach to two loss queuing systems under specific assumptions. With the deployment of server rate optimization we reach Nash equilibrium points. We also provide some analytical derivations and validate them using simulations. In our second case, we deal with one loss system with an uncapacitated queue involving quasi birth death process. We find the steady state probabilities employing two different computation techniques and calculate the expected profit for each queue in the system.
Risk assessment of autonomous vehicle using Markov decision process
Autonomous decision making of intelligent vehicle is one of the most critical and challenging module due to the fact that traffic of real world is uncertain, complex, continuous and vehicles interact with each other. In this thesis, a decision making based on reinforcement learning algorithms is proposed to represent ego vehicle behaviors interacting with the stochastic behaviors of the environmental vehicles in highway traffic. The presented solver algorithms are formulated as Markov Decision Process (MDP) for autonomous vehicle problems. Proposed algorithms are implemented in a simulation environment so that they are tested and analyzed with different scenarios. Then, efficiency of different implemented algorithms are compared based on speci fed criteria. The simulation results of tested scenarios show that ego car is capable of lane change and accelerate or decelerate in order to perform safe driving without any collision with other cars which have uncertain behavior in highway.
An application of inventory pooling / distribution models in glass industry
Supply chain is a concept in which industrial engineering applications are very important in decision making processes by requiring systematic perspective, scientific approach and correct methodology. Supply chain concept covers a wide range of processes, from raw material procurement from suppliers to product delivery to customers. Inventory management, which is one of the subheadings of the supply chain, has both operational and financial consequences. While raw material inventory levels are kept high by the management in order to reduce the risk of being stockout, this reduces the net working capital due to the decrease in the inventory turnover rate of the companies. In this thesis, the benefits and costs arising from the coordination between the factories of a company operating in different locations in case of inventory pooling and selection of appropriate suppliers are evaluated. For this purpose, data set obtained from the company operating in the glass sector is used to provide input to the established mathematical model. The model achieves optimum results that are consistent with the supplier's capacity constraint by minimizing fixed costs, distribution costs, holding costs and purchasing costs while making supplier selection and warehouse location assignments. The nonlinear model, which emerges because the safety stock level calculation contains square root expression, is linearized by using piecewise linear function. The model results are compared under the scenarios of single warehouse use, the optimum number of warehouse use determined by the model and the use of all warehouses. On the other hand, since varying sales prices and profitability rates of suppliers create inequality while comparing scenarios, these factors are eliminated from the model and what-if study is performed. As a result, the decrease in inventory levels reduces the inventory holding cost and fixed cost, thus provides a cost advantage in spite of increasing in transportation cost by not using all available warehouses which are close to factories. This shows how important it is to ensure inter-location coordination in companies operating in multiple locations.
Planning bundling and marketing efforts for a digital distribution platform
The literature on supply chain coordination is extensive, and there are numerous methods for coordinating these systems. There may be various types of associated costs in these systems related to production, quality, or sales effort, and various approaches and extensions may be integrated to such systems. Product bundling is one of the common practices that supply chain members can use to expand the market. Our study combines supply chain coordination and product bundling by incorporating marketing effort. While application developers (ADs) decide on price and marketing effort, the distribution platform (DP) decides whether to bundle developers' products and the commission rate he will charge them. This thesis adds to the literature on optimal platform bundling strategy when producers have the option to increase demand with marketing efforts. When the distribution platform implements the bundling approach, how application developers set their prices when the distribution platform offers a bundle option, and the effects of bundling on supply chain members are just a few of the questions we try to address. As a result of this study, it is shown that DP's optimal strategy on product bundling is dependent on the price ADs would announce. Secondly, ADs optimal pricing decisions may depend on the presence of bundling option provided to DP when there is a constraint on commission rate. Thirdly, it is shown that while the existence of the bundling option always benefits the distribution platform, it might benefit or harm application developers depending on the model settings. We also find that the total supply chain revenue is higher in the centralized supply chain compared to the decentralized supply chain.
Analysis of inventory policies under estimation of demand parameters
The aim of this thesis is to analyze the estimation performance for different characteristics of data in different supply chain systems. In this study, we analyze two different systems. The first section of this thesis covers the parameter estimation problem of a single retailer who observes price-dependent uncertain demand in additive form. To estimate the parameters of the demand, we use linear regression. Additionally, we jointly optimize the price and order quantity. We present a numerical analysis where parameter estimations are calculated for different experiment sets. We analyze the estimations for the number of observations, coefficient of variation, price, and profit margin. We show that as the number of observations increases and the coefficient of variation decreases, the accuracy of our estimations increases. The real optimal price of the data sets does not indicate a pattern of improvement and the profit margin of the product is observed to have no significant effect on the estimation performance. In the second section, we introduce a two-echelon system observing identical and non-identical exponentially distributed demand rates. The warehouse makes the allocation decision of its limited inventory based on the past demand data of retailers. The capacity constraint of the warehouse creates a conditional expected profit function. To find the expected profit, we propose a simulation method where the inventory allocation is repeated multiple times. To evaluate the performance of our method, we introduce experiment sets. Results show that for the systems with identical and non-identical demand rates, the percent loss in expected profit is smaller when the system utilization and the profit margin are larger. For the latter system, we analyze the effect of the ratio between the demand rates of two retailers. The results show no significant effect of this ratio on the percent loss in expected profit.
Natural gas storage valuation using deep reinforcement learning
Energy trading involves physically handling energy commodities such as oil, natural gas, and coal. These trading activities are enabled by a complex global network of energy conversion assets such as pipelines, refineries, and storage facilities. The profitability of energy trading relies on efficiently managing these assets' operational capacity constraints. The competitive management of these assets is called merchant operations. Energy merchant companies acquire conversion assets to support their trading activities. Therefore, the proper valuation of these assets is crucial. In practice, energy merchant companies rely on heuristic methods that produce deterministic operating policies. Deterministic policies are unable to capture the so-called embedded optionality in the energy conversion assets. In this thesis, we tackle the problem of natural gas storage valuation using a novel deep reinforcement learning (DRL) algorithm called soft actor-critic (SAC). SAC utilizes entropy regularization to improve policy exploration and stability. Our results show that SAC has learned an effective operating policy during training while other state-of-the-art DRL algorithms couldn't do so in our problem.
Analysis of a fork-join system with strategic customers
In real life, fork-join queuing systems can be employed to model a variety of systems including service systems, healthcare, project management, data processing, and manufacturing systems. In this thesis, we investigate the effect of the various information levels on the performance measures of a two-server fork-join queuing system with strategic customers. Strategic customers receive the information provided to them and make joining/balking decisions using this information upon their arrival. Customers arrive at the system as a pair of two, each of which is directed to one of the parallel servers if they decide to join the system. After the completion of the service, customers wait for each other and depart from the system together. %Therefore, the total time spent in the system by a pair is determined by the maximum of the two service durations. In this study, we examine three types of information structures: fully observable, partially observable, and unobservable. The level of the information provided to customers plays a significant role in their joining decision and therefore it impacts the performance of the entire system. In order to understand this impact, a set of performance measures are considered which are average utility, effective arrival (joining) rate, and expected waiting time of the customers. We derive the necessary mathematical expressions to calculate the performance measures of a two-server fork-join system for all three information structures. Moreover, a set of numerical experiments are performed with different values of the system parameters. The results show that additional information always increases the average utility of the customers. In most of the cases, the joining rate of the customers decreases as more information is provided to them. We identify the cases in which more information generates a higher effective arrival rate. Additionally, we provide a comprehensive analysis on the results of the experiments.
Analysis of production and service systems under price-dependent intermittent demand
This thesis investigates an all-or-nothing inventory and pricing problem, considering price-dependent and price and yield-dependent intermittent demand scenarios faced by a retailer. The study aims to determine the existence of a threshold policy, optimizing ordering decision and pricing strategies to maximize profit. In price-dependent intermittent demand, we explore a simple demand occurrence probability function of price with various demand size distributions. The focus is on the lost sales model due to its tractability in proving the threshold policy. In addition to the single period, a dynamic programming model for a two-period scenario is developed. Numerical analyses confirm the existence of a threshold policy, with observations indicating sensitivity to holding and unit variable costs. The thesis extends the model to incorporate yield uncertainty, introducing random yield realization and modifying the demand occurrence probability function accordingly. Responsive and unresponsive pricing schemes are explored for both single and two-period, and numerical analyses reveal threshold policies under both schemes. When the expected profit is compared, responsive pricing performs better than unresponsive pricing. This research contributes insights into the interplay of pricing, inventory, and uncertainty, providing practical applications for retail decision-making.
Portfolio optimization with sentiment analysis
Quantitative finance practices focus on developing solutions to financial markets using mathematics, statistics, and computational techniques. The objective is to maximize the return while minimizing the risk of trading. To achieve that, an optimized portfolio that is balanced between return and risk must be constructed. After the optimal portfolio is created, price predictions have to be made. Sentiment analysis is a machine learning technique to understand the sentiment in a text. It is one of the most commonly used methods to predict price direction of financial instruments. After the prediction process, portfolio is optimized using the Markowitz model developed by Harry Markowitz in 1952 by selecting a group of financial instruments. In this research, these methodologies are used for predicting the financial price and optimizing the portfolio respectively. To develop a sentiment analysis framework, words in texts must be represented as numeric values. Bag of words is a method that converts text into numeric values by using word frequency and counts. Another useful technique is word embeddings, that is learning the distributed representations of words. Several machine learning algorithms such as Naive Bayes, probabilistic machine learning model based on bayessian theorem, support vector machines, finding optimum hyperplane to maximize distances between data points which belongs to different classes, and recurrent neural network model, a neural network architecture that learns sequential patterns in data with the help of long short term memory are used. The text data is gathered from X (formerly Twitter) and the daily price data of the components of bist100 index of Borsa Istanbul A.S ̧. are employed to solve portfolio optimization problem. Our experiments show that developing a sentiment model on X data to have an optimized portfolio is useful to increase the return while minimizing the risk.