Empirical comparison of time series forecasting strategies : forecasting the Baltic p1a spot price using gradient boosting and different strategies for multi-step time series

Abstract

This NHH master thesis researches methodologies for forecasting a financial time series, the Baltic Dry P1A spot price, one week and one month ahead. The methods researched are four different strategies for time series prediction. The first is by fitting the future timestep directly based on information about today. The second is a recursive strategy, which iterates a one-step ahead prediction model. Third, a rectify implementation that corrects bias from a recursive model, by training on the residuals. Last, a direct recursive approach that fits each timestep directly with previous predictions as an added variable. Our research finds that the Direct and Direct Recursive (DirRec) strategy is the most accurate for both long and short forecast horizons. This performance is consistent when testing on an independent test set. We cannot exclude that the Direct Recursive strategy could perform better than Direct, as we get differing results when performing experiments with fewer variables. An important trade-off is that the Direct Recursive strategy is substantially more computationally heavy than the Direct strategy.