Deep Learning for Financial Time Series Forecast Fusion and Optimal Portfolio Rebalancing

Abstract

Portfolio selection is complicated by the difficulty of forecasting financial time series and the sensitivity of portfolio optimisers to forecasting errors. To address these issues, a portfolio management model is proposed that makes use of Deep Learning Models for weekly financial time series forecasting of returns. Our model uses a late fusion of an ensemble of forecast models and modifies the standard mean-variance optimiser to account for transaction costs, making it suitable for multi-period trading. Our empirical results show that our portfolio management tool outperforms the equally-weighted portfolio benchmark and the buy-and-hold strategy, using both Long Short-Term Memory and Gated Recurrent Unit forecasts. Although the portfolios are profitable, they are also sub-optimal in terms of their risk to reward ratio. Therefore, greater forecasting accuracy is necessary to construct truly optimal portfolios.