A right kind of wrong: European equity market forecasting with custom feature engineering and loss functions

Abstract

This study makes time-series-based predictions on future returns of the STOXX Europe 600 and the German DAX by adopting (in addition to a lagged and transformed version of the target series) a diversified set of predictors. Feature engineering expands further — from the initial raw group of variables, to extract knowledge of market conditions and demand for hedging. A penalisation factor is introduced with loss functions to learn a model from neural networks, in order to adapt a traditional machine learning regression framework to solve the equity forecasting problems in question. Architectures based on convolutional neural network are proposed, treating the obtained feature map similarly to an image. Experiments over different time periods demonstrate that trading strategies derived from the forecasts are more profitable than models based on efficient market assumptions. The temporal, non-stationary structure of financial data has a significant impact on the out of sample success of any model. It thus can be seen that different architectures exhibit different resilience to changing market conditions.