Within each company, the managerial area constantly needs to monitor processes and make important strategic decisions to plan the future. The decision-making process is supported by data that monitor those aspects that are crucial for the business, including sales trends. In this context, it is more and more important to have forecasts estimating how many sales will take place in the next days, weeks, and months in a reliable way. This task is called sales forecasting and is highly complex due to the influence of internal and external factors. The thesis project presented has been developed in collaboration with Amplifon S.p.A., a world leader company in the hearing aid sector interested in making predictions regarding the number of units that will be sold in different time windows in six countries where the company operates. Recently, artificial neural networks and other machine learning techniques have achieved good results when used to tackle the task of sales forecasting. However, the problem with such complex models is that the obtained results are difficult to interpret, i.e., they are black box models, and may lead the managerial area to distrust the predictions after inconsistencies or poor performances. The aim of this thesis is to predict Amplifon's sales by modeling, through a white box approach, the business and the dynamics that lead the customers to advance in the sales funnel, i.e., simulate the path of states that each customer goes through from the first visit to an Amplifon store until the moment she/he performs a purchase. In this thesis, we develop a novel model, namely the Markov Mesh, which proposes to carry out simulations in which the customers, whose behavior has been learned, move in a sequence of Markov Chains (each one models a sales funnel) and the final prediction corresponds to the number of passages through states representing the sale of hearing aids. These simulations, to provide justifiable predictions, can be analyzed a posteriori to understand which dynamics occurred as expected and where unexpected behaviors were found. Finally, the results obtained by the Markov Mesh will be compared with those produced by black box models currently used by Amplifon.
All'interno di ogni azienda l'area manageriale ha continuamente bisogno di monitorare i processi e prendere importanti decisioni strategiche per programmare il futuro. Solitamente, il processo decisionale è supportato da statistiche che monitorano aspetti determinanti per il business tra cui l'andamento delle vendite. In questo contesto risulta essere sempre più importante disporre di previsioni in grado di stimare con la maggiore accuratezza possibile quante vendite avverranno nei prossimi giorni, settimane e mesi. Tale operazione prende il nome di sales forecasting e risulta essere un compito molto complesso a causa dell'influenza di fattori interni ed esterni. Il progetto presentato in questa tesi, sviluppato in collaborazione con Amplifon S.p.A., azienda leader a livello mondiale nel settore degli apparecchi acustici, si pone l'obiettivo di avere predizioni riguardanti il numero di unità che sarà in grado di vendere in diversi archi temporali in sei dei paesi in cui l'azienda opera. Nel recente passato sono stati ottenuti buoni risultati nel fronteggiare il problema del sales forecasting grazie all'utilizzo di reti neurali e altre tecniche di machine learning. Tuttavia, il problema di modelli così complessi è che risultano essere poco interpretabili, essendo i cosiddetti modelli black box, e possono portare l'area manageriale a non fidarsi delle predizioni nel caso in cui si verifichino inconsistenze nelle predizioni o pessime prestazioni. L'obiettivo di questa tesi è quello di prevedere le vendite di Amplifon modellando, tramite un approccio white box, il business e le dinamiche che portano i suoi clienti ad avanzare nel sales funnel, ovvero simulando il percorso di stati che ciascun cliente attraversa dal primo ingresso in un negozio di Amplifon fino al momento in cui effettua un acquisto. Il modello sviluppato, detto Markov Mesh, si propone di effettuare simulazioni in cui i clienti, di cui è stato appreso il comportamento, si muovono in una sequenza di Markov Chains (in cui ciascuna modella un sales funnel) e la predizione finale corrisponde al numero di attraversamenti di stati rappresentanti la vendita di apparecchi acustici. Tali simulazioni permettono di essere analizzate a posteriori per capire quali dinamiche si sono realizzate come previsto e dove invece si sono riscontrati dei comportamenti inattesi. I risultati ottenuti dalla Markov Mesh saranno confrontati con quelli prodotti da modelli black box attualmente utilizzati da Amplifon.
Sales funnel simulation and sales forecasting with Markov chains
Fontana, Fabio
2020/2021
Abstract
Within each company, the managerial area constantly needs to monitor processes and make important strategic decisions to plan the future. The decision-making process is supported by data that monitor those aspects that are crucial for the business, including sales trends. In this context, it is more and more important to have forecasts estimating how many sales will take place in the next days, weeks, and months in a reliable way. This task is called sales forecasting and is highly complex due to the influence of internal and external factors. The thesis project presented has been developed in collaboration with Amplifon S.p.A., a world leader company in the hearing aid sector interested in making predictions regarding the number of units that will be sold in different time windows in six countries where the company operates. Recently, artificial neural networks and other machine learning techniques have achieved good results when used to tackle the task of sales forecasting. However, the problem with such complex models is that the obtained results are difficult to interpret, i.e., they are black box models, and may lead the managerial area to distrust the predictions after inconsistencies or poor performances. The aim of this thesis is to predict Amplifon's sales by modeling, through a white box approach, the business and the dynamics that lead the customers to advance in the sales funnel, i.e., simulate the path of states that each customer goes through from the first visit to an Amplifon store until the moment she/he performs a purchase. In this thesis, we develop a novel model, namely the Markov Mesh, which proposes to carry out simulations in which the customers, whose behavior has been learned, move in a sequence of Markov Chains (each one models a sales funnel) and the final prediction corresponds to the number of passages through states representing the sale of hearing aids. These simulations, to provide justifiable predictions, can be analyzed a posteriori to understand which dynamics occurred as expected and where unexpected behaviors were found. Finally, the results obtained by the Markov Mesh will be compared with those produced by black box models currently used by Amplifon.File | Dimensione | Formato | |
---|---|---|---|
2021_12_Fontana.pdf
non accessibile
Descrizione: Versione finale tesi e summary
Dimensione
2.44 MB
Formato
Adobe PDF
|
2.44 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/183574