Prediction of Parallel Time Series with Machine Learning

Magistrsko delo obravnava napovedovanje vzporednih časovnih vrst s strojnim učenjem. Vzporedne časovne vrste so časovne vrste, katerih vrednosti se spreminjajo v času ob enakih časovnih intervalih, kot je npr. ura ali dan istočasno za vse časovne vrste. Primer take časovne vrste so borzni tečaji, kjer se za vsak vrednostni papir posebej tvori ena časovna vrsta vzporedno s časovnimi vrstami ostalih vrednostnih papirjev. Doprinos magistrske naloge je v novem kombiniranem algoritmu za napovedovanje vzporednih časovnih vrst, ki vsebuje genetski algoritem in nelinearno regresijo. Genetski algoritem je uporabljen za iskanje sita in nelinearne funkcije, ki opisujeta model. Za izračunanje neznanih koeficientov funkcije se uporablja numerična metoda nelinearne regresije. Novo predlagani algoritem je primerljiv glede točnosti in mere dobička z obstoječimi algoritma strojnega učenja. Prednost je da za vhodne podatke ne rabi posebnih predobdelav podatkov, dokler so le ti polni. Druga prednost je tudi, da ponuja razlago, kako so podatki odvisni med sabo. Slabost algoritma pa je njegova časovna potratnost, ki smo se ji v delu delno izognili s paralelizacijo.This master\u27s thesis deals with the prediction of parallel time series with the use of machine learning. Parallel time series are time series whose values change over time at equal time intervals, such as hour or day simultaneously for all time series. An example of this type of time series are stock exchange rates, where for each security a single time series is created parallel to the time series of other securities. The contribution of this master\u27s thesis is a new combined algorithm for predicting parallel time series that includes a genetic algorithm and non-linear regression. The genetic algorithm is used to find the sieve and for non-linear functions that describe the model. The numerical method of nonlinear regression is used to calculate unknown function coefficients. The new proposed algorithm is comparable in terms of accuracy and profit margin with existing machine learning algorithms. The advantage of the algorithm is that it does not need specific data preprocessing for input data as long as data are complete. Another advantage is that it offers an explanation on how data depend on one another. The downside is that the algorithm is time consuming, which was partly avoided with parallelism