Browsing by Subject "Markov Chain Monte Carlo"

Now showing 1 - 3 of 3

Open Access
Learning price promotion effects on recurring sell-in purchases from simulated store level sales data
(2021-06) Keşrit, Pelin
When a product is put on promotion to increase its sales, this causes a decrease in the sales of another product in the same product group. This phenomenon happens usually when the promoted product is a substitute for the other product. In this study, we focus on the wholesaler’s revenue maximization problem over the given planning horizon. For this purpose, we constructed a Bayesian hierarchical model for the order quantities observed in the store level data for substitutable products. Order quantities are assumed to have Poisson distributions whose means depend on season, prices and previously ordered quantities for all products in the same group. The customers are assumed to have diﬀerent price sensitivities, and consumption rates implicit in their historical order quantities. Using a hybrid of diﬀerent Markov Chain Monte Carlo methods, we update model parameter posterior distributions and predict each retailer’s order quantities in the future. We veriﬁed on simulated sales data that the MCMC methods work.
Open Access
Meal participation prediction with bayesian hierarchical models
(2021-12) Kof, Aleyna
Forecasting sales in the catering industry helps authorities to organize daily transactions eﬃciently to prevent both waste and business loss. In this study, we focused on predicting meal sales in Bilintur Catering Centre with the dataset which is collected through ﬁve academic years. To forecast the meal sales, we constructed two Bayesian hierarchical models. The ﬁrst model does not diﬀerentiate eﬀects of predictors in diﬀerent academic years, while the second does. We derived the full conditional distributions and employed Gibbs sampling in an extensive MCMC study. We tested two models along with a benchmark multiple regression model on the held-out academic year. We concluded that multiple regression and ﬁrst model provide more accurate results.
Open Access
Missed flight cover design
(2019-07) Çelik, Beyza
Missed flight cover is an option with a price and validity period and is a source of ancillary revenues for the airline companies and helps passengers, who missed their flights, resume their journeys at reduced costs. We study optimal price and validity period of this option to allow a passenger to use missed flight fare towards the purchase of a future airline ticket. Our objective is to maximize the expected ancillary revenues of the airline. The possible actions of passengers are described with a probabilistic graphical model. Within that model, passenger's decision to buy the option and to resume the journey after a missed flight are described with separate hierarchical Bayesian mixed logit regression models. To estimate the parameters of those mixed logit models, an individualized Bayesian choice-based conjoint experiment is designed. In this experiment, each choice set is optimally picked so as to maximize the expected Kullback-Leibler divergence between subsequent posterior distributions of individualized part-worths. The posterior distributions of unknown model parameters, particularly, individualized part-worths, are calculated with a hybrid Markov Chain Monte Carlo (MCMC) algorithm. We developed an R-Shiny online survey web application for six di erent individualized choice experiments (buy or not buy an option for leisure and business travel, resume or not resume a missed leisure or business flight with or without an option) and collected responses of over 300 individuals. Using the MCMC samples of individual part-worths from their posterior distributions, we simulated the market. We searched and found an option design that maximized the average net revenue of the airline over the simulated runs of the market.