Study On Forecasting BFI Based On Wavelet Analysis And Neural Network

Dry-bulk shipping market has the largest risk and the most violent fluctuation in the international shipping market. Baltic freight index (BFI) Ts considered as barometer of dry-bulk shipping market which can guide effectively many ship-owners and freighters to venture investment. BFI fluctuates violently and gets out of trend for being influenced by many factors, which bring huge risk and opportunity for operators in market. Just going on that point, the thesis combines qualitative and quantitative analysis, mainly adopts wavelet analysis and neural network to search for inherent laws of index and forecasting index effectively.Dry-bulk shipping market's structure, character, especially cargo requirement and transport capability supply have been deeply analyzed in the thesis. Market requirement and supply are the two main factors to effect index. In recent years requirement in dry-bulk segment market, development and transport capacity of every type of ships and the influential factors on requirement and supply also have been analyzed, which lay the groundwork to analyze dry-bulk shipping market macroscopically and index trend. Then BFI's generation, development, algorithmic method and historical march have been analyzed, and its character also has been completely introduced and analyzed in the thesis.Wavelet analysis and neural network are combined to forecast BFI. BP neural network has huge capacity to non linearity simulate, self-organize and self-study, can grasp the inherent laws of things well. But many irregular and whooping factors are included in the influence factors on BFI, which have a strong impact on neural network to self-study as noise. Wavelet analysis is adopted to de-noise BFI and get rid of noise, so neural network will simulate and forecast better.From the forecasted results, that thought can realize better to grasp the development trend of BFI, the errors between forecasted and real results are held under 4%. Those results are compared with the results only forecasted by neuralnetwork not de-noising in the thesis. The conclusion is that the former can hold the index's inherent laws better, deviation degree between forecasted and real results is lower, index fluctuates more gently, and error rate is also lower.