Time series analysis and forecasting are two increasingly important fields of study which are attracting more and more attention in recent years. This is due to the massive adoption of time series data in real world scenarios, like in healthcare, finance, sales and weather control. Indeed, analyzing time series data, it is possible to acquire useful knowledge to better understand a given phenomenon and to support decision making. A practical example is represented by the Covid-19 pandemic, where time series data (e.g. daily cases, daily intensive care) help governments and health institutions to monitor the emergency and make decision to limit the spread of the virus. However, several techniques of time series analysis and forecasting assume the data to be stationary to proper perform their tasks. Nevertheless, this assumption is generally not respected in time series representing real phenomena. Indeed, these series are often characterized by non-stationary dynamics like trends or seasonality in the data. The purpose of the thesis is to present a tool able to mitigate the problem of non-stationarity in time series data. The tool is named MATIS (Making A Time series Stationary) and is able to manage non-stationary time series, automatically detecting and removing the non-stationary dynamics within the data and thus obtaining a stationary version of the series. Several tests were performed to assess the accuracy of the proposed solution in detecting and removing non-stationary dynamics in time series. Moreover, specific tests were conducted to verify whether MATIS has a positive impact on time series forecasting and correlation analysis. The experimental results obtained suggest that the proposed tool can be an helpful solution when working with non-stationary time series.
L'analisi e la predizione delle serie temporali sono due campi di studio che stanno guadagnando sempre più importanza negli ultimi anni. Questo è dovuto al massiccio utilizzo dei dati delle serie temporali che viene fatto in diversi settori lavorativi, come nel settore sanitario, finanziario, nelle vendite e nel controllo climatico. Infatti, analizzando le serie, è possibile acquisire informazioni utili per comprendere meglio un certo fenomeno e per supportare il processo decisionale. Un esempio pratico è rappresentato dalla pandemia di Covid-19 dove i dati delle serie temporali aiutano i governi e le istituzioni sanitarie a prendere decisioni per limitare la diffusione del virus. Diverse tecniche di analisi e predizione delle serie temporali assumono la condizione di stazionarietà della serie per poter funzionare correttamente. Tuttavia questa assunzione viene spesso non rispettata dalle serie rappresentanti fenomeni del mondo reale. Queste serie sono infatti spesso caratterizzate da dinamiche non stazionarie, come la presenza di trend o stagionalità. Lo scopo di questa tesi è presentare uno strumento capace di mitigare il problema di non stazionarietà all'interno delle serie temporali. Lo strumento è chiamato MATIS (Making A Time series Stationary) ed è in grado di gestire serie non stazionarie, identificando e rimuovendo in modo automatico le dinamiche non stazionarie presenti ed ottenendo una sua versione stazionaria. Diversi test sono stati effettuati per verificare l'accuratezza della soluzione proposta nell'individuare e rimuovere le dinamiche non stazionarie all'interno delle serie. Inoltre, test specifici sono stati condotti per verificare se MATIS abbia un impatto positivo sulla predizione ed analisi delle serie temporali. I risultati sperimentali ottenuti suggeriscono che lo strumento proposto possa essere un'utile soluzione per lavorare con serie temporali non stazionarie.
Stationarity is (not) all you need : the MATIS algorithm
Tomacelli, Francesco
2020/2021
Abstract
Time series analysis and forecasting are two increasingly important fields of study which are attracting more and more attention in recent years. This is due to the massive adoption of time series data in real world scenarios, like in healthcare, finance, sales and weather control. Indeed, analyzing time series data, it is possible to acquire useful knowledge to better understand a given phenomenon and to support decision making. A practical example is represented by the Covid-19 pandemic, where time series data (e.g. daily cases, daily intensive care) help governments and health institutions to monitor the emergency and make decision to limit the spread of the virus. However, several techniques of time series analysis and forecasting assume the data to be stationary to proper perform their tasks. Nevertheless, this assumption is generally not respected in time series representing real phenomena. Indeed, these series are often characterized by non-stationary dynamics like trends or seasonality in the data. The purpose of the thesis is to present a tool able to mitigate the problem of non-stationarity in time series data. The tool is named MATIS (Making A Time series Stationary) and is able to manage non-stationary time series, automatically detecting and removing the non-stationary dynamics within the data and thus obtaining a stationary version of the series. Several tests were performed to assess the accuracy of the proposed solution in detecting and removing non-stationary dynamics in time series. Moreover, specific tests were conducted to verify whether MATIS has a positive impact on time series forecasting and correlation analysis. The experimental results obtained suggest that the proposed tool can be an helpful solution when working with non-stationary time series.| File | Dimensione | Formato | |
|---|---|---|---|
|
2021_10_Tomacelli.pdf
accessibile in internet solo dagli utenti autorizzati
Descrizione: Tesi Laurea Magistrale
Dimensione
3.07 MB
Formato
Adobe PDF
|
3.07 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/10589/179286