Application of Wavenet to financial times series prediction

Abstract

This thesis explores the application of the WaveNet model utilizing dilated causal convolutions, originally designed for text-to-speech synthesis on univariate time series. Here it is adapted to predicting on multivariate financial time series focusing on 43 commodities consisting of currencies, bonds, indexes, soft commodities, metal, grains, energy and live stock. The study trains, and evaluates the models on commodities from Pinnacle data corp's CLC database, with the loss function of mean squared error and model performance is evaluated on a scorecard of metrics like accuracy, edge, noise, and calibration ratio. The results reveal varying performance across commodities, with heating oil demonstrating the highest edge of 0.005 and accuracy of 53% and rbob gasoline having the second highest edge with 0.001 and accuracy of 52.2%. Outcomes demonstrate that around half the models had a higher than 50% accuracy in predicting the daily movement and about half, below. New insight into adapting WaveNet to multivariate time-series prediction is found, and discussion highlights the potential of hyper-parameter tuning and suggest further exploration of advanced models and alternative data-sets for enhanced predictions. The findings underscore further investigation into deep learning in financial forecasting, with implications for trading strategies and future model refinement, but does not contradicts the efficient market hypothesis. A comparison is made on the results of one commodity, Sugar, to a GARCH(1,1) model.

This thesis explores the application of the WaveNet model utilizing dilated causal convolutions, originally designed for text-to-speech synthesis on univariate time series. Here it is adapted to predicting on multivariate financial time series focusing on 43 commodities consisting of currencies, bonds, indexes, soft commodities, metal, grains, energy and live stock. The study trains, and evaluates the models on commodities from Pinnacle data corp's CLC database, with the loss function of mean squared error and model performance is evaluated on a scorecard of metrics like accuracy, edge, noise, and calibration ratio. The results reveal varying performance across commodities, with heating oil demonstrating the highest edge of 0.005 and accuracy of 53% and rbob gasoline having the second highest edge with 0.001 and accuracy of 52.2%. Outcomes demonstrate that around half the models had a higher than 50% accuracy in predicting the daily movement and about half, below. New insight into adapting WaveNet to multivariate time-series prediction is found, and discussion highlights the potential of hyper-parameter tuning and suggest further exploration of advanced models and alternative data-sets for enhanced predictions. The findings underscore further investigation into deep learning in financial forecasting, with implications for trading strategies and future model refinement, but does not contradicts the efficient market hypothesis. A comparison is made on the results of one commodity, Sugar, to a GARCH(1,1) model.