An Analysis of UK Natural Gas Prices Comparing Regression, Neural Nets and Gradient Boosting Techniques
Master thesis
Permanent lenke
https://hdl.handle.net/11250/3101224Utgivelsesdato
2023Metadata
Vis full innførselSamlinger
Sammendrag
I denne masteroppgaven ble det utført en komparativ analyse av de tre ulike modelleringsmetodene: multippel lineær regresjon, gradient boosting metoden XGBoost og detnevrale nettverket Long Short-Term Memory (LSTM). For ˚a sammenligne modellene bletidsserie data for naturgassprisen i UK brukt.M˚alet med denne masteroppgaven var ˚a utforske i hvilken grad de ulike modelleringsmetodene klarte ˚a konstruere sammenhenger mellom variabler, som p˚avirker naturgassprisen,og vurdere deres prediksjonsytelse for en svært volatil handelsvare som gassprisen. Dissetre modellene representerer forskjellige grener av modellering som tradisjonell statistiskmodellering, maskinlæring og nevrale nettverk.To ulike perioder ble valgt for ˚a vurdere modellene: en periode under en stabil markedssituasjon mellom 2018 og begynnelsen av 2020, og en annen periode under en markedskrisesom startet med gassleverandørkrisen i Storbritannia i 2021 og fortsatte med Russlandsinvasjon av Ukraina. Dataene dekker perioden fra begynnelsen av 2009 til slutten av2022.I den ”normale” perioden presterte alle tre modellene tilfredsstillende med hensyn tilprediksjon, der multiple lineær regresjonsmodell viste en noe svakere ytelse enn de toandre. LSTM-modellen presterte marginalt bedre enn XGBoost. I den svært volatileperioden lyktes regresjonsmodellen med ˚a fange opp pris-dynamikken og identifiserte toav de tre store pris-hoppene.For den gjennomsnittlige prediksjonsnøyaktigheten var LSTM-modellen best, med XGBoostmodellen like bak og sist, regresjonsmodellen. B˚ade LSTM og XGBoost hadde problemermed ˚a ekstrapolere de bratte prisstigningene, men viste fortsatt en god tilpasningsevne. This master thesis presents a comparative analysis of three different modeling methods:multiple linear regression, the gradient boosting technique XGBoost, and the neural network Long Short-Term Memory (LSTM). To compare the models, the financial time seriesof the natural gas price in the UK was used.The main goal of this investigation was to explore the effectiveness of different modelingapproaches in capturing the complex relationships between variables that impact thenatural gas price and assess their predictive performance on a highly volatile commoditysuch as the gas price. These models represent distinct branches of modeling, includingtraditional statistical modeling, machine learning, and neural networks.Two different periods were selected to assess the models’ efficacy, one period during astable market between 2018 and the beginning of 2020, and the other during a marketcrisis that started with the 2021 natural gas supplier crisis in the UK and continued withthe Russian invasion of Ukraine. The data covered the period from the beginning of 2009to the end of 2022.During the ”normal” period, all three models performed adequately in terms of prediction,with the multiple linear regression model showing a slightly weaker performance than theother two. The LSTM model marginally outperformed XGBoost. During the highlyvolatile period, the multiple linear regression model demonstrated success in capturingthe price dynamics and did successfully identify two significant price surges, out of thethree observed.In terms of overall prediction accuracy, the LSTM model exhibited the best performance,with the XGBoost model a close second, and the multiple linear regression model exhibiteda comparatively lower performance. However, both LSTM and XGBoost struggled toextrapolate the price surges but still demonstrated a strong overall fit.