We investigate the benefits of feature selection, nonlinear modelling and online learning with forecasting in financial time series. We consider the sequential and continual learning sub-genres of online learning. Through empirical experimentation, which involves long term forecasting in daily sampled cross-asset futures, and short term forecasting in minutely sampled cash currency pairs, we find that the online learning techniques outperform the offline learning ones. We also find that, in the subset of models we use, sequential learning in time with online Ridge regression, provides the best next step ahead forecasts, and continual learning with an online radial basis function network, provides the best multi-step ahead forecasts. We combine the benefits of both in a precision weighted ensemble of the forecast errors and find superior forecast performance overall.

翻译：我们调查地物选择、非线性建模和在线学习以及财务时间序列预测的好处。我们考虑在线学习的连续和持续学习子模式。我们通过实验实验(包括对每日抽样交叉资产期进行长期预测,对每分钟抽样现金货币对子进行短期预测 ) 发现在线学习技术优于离线学习技术。 我们还发现,在所使用模型的子集中,与在线山脊回归同步的连续学习,提供了最佳的下一步预测,与在线半径基功能网络的持续学习,提供了最佳的多步预报。 我们把两者的效益结合到预测误差的精确加权组合中,并找到高超的预测总体性能。