Improved sales forecasts for individual products in retail stores can have a positive effect both environmentally and economically. Historically these forecasts have been done through a combination of statistical measurements and experience. However, with the increased computational power available in modern computers, there has been an interest in applying machine learning for this problem. The aim of this thesis was to utilize two years of sales data, yearly calendar events, and weather data to investigate which machine learning method could forecast sales the best. The investigated methods were XGBoost, ARIMAX, LSTM, and Facebook Prophet. Overall the XGBoost and LSTM models performed the best and had a lower mean absolute value and symmetric mean percentage absolute error compared to the other models. However, Facebook Prophet performed the best in regards to root mean squared error and mean absolute error during the holiday season, indicating that Facebook Prophet was the best model for the holidays. The LSTM model could however quickly adapt during the holiday season improved the performance. Furthermore, the inclusion of weather did not improve the models significantly, and in some cases, the results were worsened. Thus, the results are inconclusive but indicate that the best model is dependent on the time period and goal of the forecast. / Förbättrade försäljningsprognoser för individuella produkter inom detaljhandeln kan leda till både en miljömässig och ekonomisk förbättring. Historiskt sett har dessa utförts genom en kombination av statistiska metoder och erfarenhet. Med den ökade beräkningskraften hos dagens datorer har intresset för att applicera maskininlärning på dessa problem ökat. Målet med detta examensarbete är därför att undersöka vilken maskininlärningsmetod som kunde prognostisera försäljning bäst. De undersökta metoderna var XGBoost, ARIMAX, LSTM och Facebook Prophet. Generellt presterade XGBoost och LSTM modellerna bäst då dem hade ett lägre mean absolute value och symmetric mean percentage absolute error jämfört med de andra modellerna. Dock, gällande root mean squared error hade Facebook Prophet bättre resultat under högtider, vilket indikerade att Facebook Prophet var den bäst lämpade modellen för att förutspå försäljningen under högtider. Dock, kunde LSTM modellen snabbt anpassa sig och förbättrade estimeringarna. Inkluderingen av väderdata i modellerna resulterade inte i några markanta förbättringar och gav i vissa fall även försämringar. Övergripande, var resultaten tvetydiga men indikerar att den bästa modellen är beroende av prognosens tidsperiod och mål.

Submitted by Automa??o e Estat?stica (sst@bczm.ufrn.br) on 2017-06-02T22:45:04Z No. of bitstreams: 1 ThiagoRodriguesDeSouza_DISSERT.pdf: 4007098 bytes, checksum: b98a621f29e991b3bc26a68f0060557f (MD5) / Approved for entry into archive by Arlan Eloi Leite Silva (eloihistoriador@yahoo.com.br) on 2017-06-09T20:35:17Z (GMT) No. of bitstreams: 1 ThiagoRodriguesDeSouza_DISSERT.pdf: 4007098 bytes, checksum: b98a621f29e991b3bc26a68f0060557f (MD5) / Made available in DSpace on 2017-06-09T20:35:17Z (GMT). No. of bitstreams: 1 ThiagoRodriguesDeSouza_DISSERT.pdf: 4007098 bytes, checksum: b98a621f29e991b3bc26a68f0060557f (MD5) Previous issue date: 2017-02-21 / O objetivo deste trabalho ? realizar um estudo comparativo com ajustes de modelos de previs?es pelo m?todo de Box-Jenkins (ARIMA) e Box-Tiao (ARIMAX) para precipita??o acumulada mensal em seis cidades do Nordeste do Brasil, sendo escolhida de acordo com a classifica??o clim?tica de K?ppen. Tendo como vari?veis ex?genas: temperaturas da superf?cie do mar do oceano Atl?ntico e Pac?fico. Em todas as s?ries de precipita??o acumulada verificou-se a presen?a do componente sazonal, al?m disso, devido ao pressuposto de vari?ncia constante e normalidade dos dados n?o serem atendida, foi aplicado na s?rie original ? transforma??o Box Cox. Atrav?s das medidas de qualidade dos ajustes dos modelos pelo m?todo ARIMA e ARIMAX, temos que o modelo ARIMAX evidenciou como o melhor ajuste aos dados em estudo, apresentando menores valores para os crit?rios de informa??o AIC, erro m?dio e erro quadr?tico m?dio. / The objective this work is realize a comparative study with adjustment of previsions models by Box-Jenkins (ARIMA) and Box-Tiao (ARIMAX) methods for monthly accumulated precipitation in six cities of Brazilian northeast, choosing the cities according with K?ppen climatic classification. We've exogenes variables: sea surface temperature of Atlantic and Pacific Ocean.In all precipitations accumulated series were observerd the presence of sazonal component, besides that, due to assumption of the constante variance and data normality isn't reached, was applied in original serie the Box Cox transformation.By the measures of quality of the models adjustments by ARIMA and ARIMAX method, we've the ARIMAX model evidencied like the better adjustment to data, showing lower values to AIC information criteria, mean error and mean square error.

The study aimed to define and investigate relevant parameters affecting manual frequency restoration reserve (mFRR) volumes of the balancing market in the Finnish price area. It also aimed to find suitable models and investigate Day-ahead prediction possibilities of mFRR volumes. Parameters related to mFRR volumes Day-ahead predictions were identified in several earlier studies where of nine parameters were investigated. The correlations between mFRR volumes and the different parameters were investigated using Spearman’s correlation. Different linear machine learning models for Day-ahead predictions of mFRR volumes were builtand tested in Python. The resulting models used for predicting mFRR volumes in Python were one ARIMAX model and one SARIMAX model. The models were validated with a walk-forward method where Day-ahead predictions were conducted monthly for one year. The accuracy of the predictions was measured by the validation parameters Mean Absolute Value, Root Mean Square Error and Median Absolute Deviation. Results from the study show that it is difficult to predict absolute activated mFRR volumes. Although, it might be possible to predict that mFRR volumes will be activated or not, up- or down regulated to some extent. One explanation of the difficulties in predicting mFRR volumes is dueto mFRR being a balancing product whose function is to regulate disturbances in the electricity grid.

This thesis focuses on modelling non-maturing deposits (NMD) and has been written in collaboration with Svenska Handelsbanken. The methodology includes regression analysis and time series analysis, with the Repo rate serving as an exogenous variable in both models. A Vasicek model is employed to generate future Repo rates, which are then used as inputs for forecasting the NMD volume. These simulated rates are then compared to forecasted Repo rates with discrete changes from an external source. The results are utilised to analyze how net interest income can vary in the case of constant volume and in the case of interest rate-dependent volume. Effective liquidity management is crucial for banks, and NMDs are an important source of funding. By using regression analysis and time series analysis, combined with the Repo rate as the exogenous variable, this thesis provides insights into the behaviour of NMD volumes, and how it is affected by the Repo rate. The models also enable the forecasting of future trends based on future Repo rates. Additionally, by using different data sets as input for future Repo rates, the behaviour of the model can be evaluated based on how well it coincides with reality. The results obtained from this analysis can also be used to compare the value and interest rate sensitivity of NMD products. In conclusion, this thesis provides an approach to modelling the NMD volumes using exogenous factors and demonstrates how this can affect the net interest income from deposit volumes. / Denna avhandling fokuserar på modellering av icke tidsbunden inlåning (non-maturing deposits, NMD) och har skrivits i samarbete med Svenska Handelsbanken. Metoden omfattar regressionsanalys och tidsserieanalys, där reporäntan fungerar som en exogen variabel i båda modellerna. En Vasicek-modell används för att generera framtida reporäntor, som sedan används som indata för att prognostisera NMD-volymen. Dessa simulerade räntor jämförs sedan med prognostiserade reporäntor med diskreta förändringar från en extern källa. Resultaten används sedan för att analysera hur räntenettot kan variera mellan fallet med konstant volym och fallet med ränteberoende volym. En effektiv likviditetshantering är avgörande för banker, och NMD:er är en viktig finansieringskälla. Genom att använda regressionsanalys och tidsserieanalys, i kombination med reporäntan som exogen variabel , ger denna avhandling värdefulla insikter i NMD-volymernas beteende och hur de påverkas av reporäntan. Modellerna gör det också möjligt att prognostisera framtida trender utifrån framtida reporäntor. Genom att använda olika datamängder som indata för de framtida reporäntorna kan modellens beteende dessutom värderas utifrån hur väl det sammanfaller med verkligheten. Resultaten från denna analys kan också användas för att jämföra NMD-produkternas värde- och räntekänslighet. Sammanfattningsvis ger denna avhandling ett tillvägagångssätt för att modellera NMD-volymerna med hjälp av exogena faktorer och visar på hur det kan påverka inlåningens räntenetto.