Bayesovi pristopi ocenjevanja dinamičnih sistemov za potrebe napovedovanja dinamike pretovora pristanišč : doktorska disertacija

slovenian

doktorska dela

2018-2019

DFA   dissertations   dynamic factor analysis   logistocs   macroeconomic indicators   MC simulation   port throughput   predictive analytics   time series

Univerza v Mariboru, Fakulteta za logistiko

Future events forecasting such as a prediction of demand is nowadays an industry standard. Maritime industry is not an exception since the quality of forecasting of a future port's throughput dynamics is the baseline for port task planning. In our work, an algorithm, which combines several fields of statistics and econometrics, is presented. Algorithm's primary goal is to provide fairly accurate future port's throughput predictions. The idea stands on modeling the output time series concerning the selected inputs. Latter are presented in the form of macroeconomic indicators, which are a priori selected from a bigger set of indicators. For this purpose, an initial data reduction of exogenous indicators has been conducted by regressing different combination of subsets of exogenous indicators randomly chosen in the Monte Carlo procedure, where the optimal set of genuinely influential indicators was chosen by observing the model's error-based criteria adopted in the multiple regression static procedure. In the next stage, the reduced set of influential indicators is aggregated into dynamic factor models using the EM algorithm and Kalman filter. Derived dynamic factors are used as inputs into the ARIMAX time series models. Complete mechanism is executed based on five-step heuristic procedure, where generating different ARIMAX candidates eventually leads to the selection of the final best candidate. The obtained ARIMAX model is robust, complies with adequate statistical and econometrics tests, and last, but not least it can provide quite accurate forecasts. The algorithm has been applied to the real port's data. Achieved ARIMAX predictions of the throughput data values were compared with the standard benchmarking models' predictions, whereas the results are promising and reveal a high level of applicability.

07.08.2023