Accurate revenue forecasting is a cornerstone of corporate financial planning and strategic decision-making. The traditional bottom-up approaches, as typically practiced for instance by EVS Broadcast Equipment S.A., depend to a large extent on managerial judgment and expert opinion, which, though valuable, tend to be subjective and susceptible to bias. This thesis presents a complementary data-driven framework, applying machine learning and factor-based modeling in forecasting corporate revenues with greater statistical rigor. Drawing on thirty years of annual data (1995–2024) from EVS, this study combines internal financial indicators, industry-specific variables, and macroeconomic factors into one complete dataset. A hierarchical modelling strategy is implemented across four stages: first, baseline machine learning models without dimensionality reduction; second, models with Principal Component Analysis embedded; third, models including Supervised PCA; and lastly, models using Generalized Iterative SPCA following the work of Giglio, Xiu, and Zhang (2023). Each model will range from linear regressions and tree-based ensembles to kernel regressors and deep learning architectures such as MLP, LSTM, GRU, and TCN; time-series cross-validation will be done for each using metrics like RMSE, MAE, R², MAPE, and SMAPE. Empirical results clearly indicate that deep learning models, specifically the Temporal Convolutional Network, outperform classical econometric and ensemble methods in modeling nonlinearities and temporal dependencies in EVS's revenue data. Dimensionality reduction via Iterative SPCA reinforces predictivity by isolating latent, economically meaningful components from a host of weak and interrelated predictors. Final forecasts for the year 2025 are augmented with Monte Carlo simulations to build 80% and 95% predictive intervals, quantifying forecast uncertainty. The overall results demonstrate that the integration of factor-based machine learning and supervised dimensionality reduction constitutes a robust, interpretable framework for corporate revenue forecasting. While this approach represents a complementary aspect to traditional managerial forecasting at EVS, it also has wider implications for mid-sized technology firms looking to embed data-driven insights into long-term strategic planning.
Una previsione accurata dei ricavi rappresenta un pilastro fondamentale della pianificazione finanziaria aziendale e del processo decisionale strategico. I metodi tradizionali di tipo bottom-up, come quelli comunemente adottati da EVS Broadcast Equipment S.A., si basano in larga misura sul giudizio dei dirigenti e sull’opinione degli esperti che, pur avendo valore, risultano spesso soggettivi e suscettibili a bias. La presente tesi propone un quadro complementare basato sui dati (data-driven), applicando tecniche di machine learning e modelli fattoriali per la previsione dei ricavi aziendali con maggiore rigore statistico. Basandosi su trent’anni di dati annuali (1995–2024) di EVS, lo studio integra indicatori finanziari interni, variabili specifiche del settore e fattori macroeconomici in un unico dataset completo. È stata implementata una strategia di modellazione gerarchica articolata in quattro fasi: (1) modelli di machine learning di base senza riduzione dimensionale; (2) modelli con Principal Component Analysis (PCA); (3) modelli con Supervised PCA; e infine (4) modelli che utilizzano la Generalized Iterative SPCA ispirata al lavoro di Giglio, Xiu e Zhang (2023). I modelli spaziano da regressioni lineari e ensemble ad alberi decisionali fino a regressori con kernel e architetture di deep learning come MLP, LSTM, GRU e TCN. Per ciascun modello è stata eseguita una time-series cross-validation utilizzando metriche quali RMSE, MAE, R², MAPE e SMAPE. I risultati empirici mostrano chiaramente che i modelli di deep learning, in particolare la Temporal Convolutional Network, superano i metodi econometrici classici e gli ensemble nel modellare le non linearità e le dipendenze temporali nei dati sui ricavi di EVS. La riduzione dimensionale tramite Iterative SPCA migliora ulteriormente la capacità predittiva, isolando componenti latenti ed economicamente significative da un insieme di predittori deboli e correlati. Le previsioni finali per il 2025 sono arricchite con simulazioni Monte Carlo per costruire intervalli predittivi all’80% e al 95%, quantificando l’incertezza delle previsioni. Nel complesso, i risultati dimostrano che l’integrazione tra machine learning basato su fattori e riduzione dimensionale supervisionata costituisce un quadro robusto e interpretabile per la previsione dei ricavi aziendali. Sebbene tale approccio rappresenti un complemento ai metodi previsionali tradizionali utilizzati da EVS, esso offre anche implicazioni più ampie per le imprese tecnologiche di medie dimensioni che intendono integrare analisi data-driven nella pianificazione strategica di lungo periodo.
Previsione dei Ricavi Aziendali tramite Machine Learning: Evidenze da EVS Broadcast Equipment S.A.
PANAHBEKHODA, EMAD
2024/2025
Abstract
Accurate revenue forecasting is a cornerstone of corporate financial planning and strategic decision-making. The traditional bottom-up approaches, as typically practiced for instance by EVS Broadcast Equipment S.A., depend to a large extent on managerial judgment and expert opinion, which, though valuable, tend to be subjective and susceptible to bias. This thesis presents a complementary data-driven framework, applying machine learning and factor-based modeling in forecasting corporate revenues with greater statistical rigor. Drawing on thirty years of annual data (1995–2024) from EVS, this study combines internal financial indicators, industry-specific variables, and macroeconomic factors into one complete dataset. A hierarchical modelling strategy is implemented across four stages: first, baseline machine learning models without dimensionality reduction; second, models with Principal Component Analysis embedded; third, models including Supervised PCA; and lastly, models using Generalized Iterative SPCA following the work of Giglio, Xiu, and Zhang (2023). Each model will range from linear regressions and tree-based ensembles to kernel regressors and deep learning architectures such as MLP, LSTM, GRU, and TCN; time-series cross-validation will be done for each using metrics like RMSE, MAE, R², MAPE, and SMAPE. Empirical results clearly indicate that deep learning models, specifically the Temporal Convolutional Network, outperform classical econometric and ensemble methods in modeling nonlinearities and temporal dependencies in EVS's revenue data. Dimensionality reduction via Iterative SPCA reinforces predictivity by isolating latent, economically meaningful components from a host of weak and interrelated predictors. Final forecasts for the year 2025 are augmented with Monte Carlo simulations to build 80% and 95% predictive intervals, quantifying forecast uncertainty. The overall results demonstrate that the integration of factor-based machine learning and supervised dimensionality reduction constitutes a robust, interpretable framework for corporate revenue forecasting. While this approach represents a complementary aspect to traditional managerial forecasting at EVS, it also has wider implications for mid-sized technology firms looking to embed data-driven insights into long-term strategic planning.| File | Dimensione | Formato | |
|---|---|---|---|
|
Emad_Panahbekhoda_MasterThesis.pdf
accesso aperto
Dimensione
1.93 MB
Formato
Adobe PDF
|
1.93 MB | Adobe PDF | Visualizza/Apri |
È consentito all'utente scaricare e condividere i documenti disponibili a testo pieno in UNITESI UNIPV nel rispetto della licenza Creative Commons del tipo CC BY NC ND.
Per maggiori informazioni e per verifiche sull'eventuale disponibilità del file scrivere a: unitesi@unipv.it.
https://hdl.handle.net/20.500.14239/31923