Capturing alternative data in practice: How can neural networks understand news sentiment?

Axel, you recently published your research “Learning deep news sentiment representations for macro-finance” in Digital Finance. What motivated you to explore this topic?

Whenever we have tried to bring together alternative sentiment data and financial time series, the biggest obstacle has always been the selection or weighting of the underlying narratives or topics. Some form of dimension reduction is required to link news collected on potentially thousands of topics to individual financial instruments. However, most of the standard statistical approaches left something to be desired. Often, too many data were discarded while other approaches produced weightings for topics irrespective of our use case context.

Given that neural networks are strong at these kinds of tasks, I was simply pursuing an idea I had about how to use neural networks to give us what we wanted. Recent software advances have made it very easy to customise and train neural networks, so it was just a question of seeing how the strengths of neural networks would carry over to our finance use case with considerably fewer data points than are typically used for training these models.

How did you approach these questions in detail?

I started the analysis using the standard dimension reduction methods like principle component analysis (PCA). However, it quickly became clear that due to the complexity of news sentiment data we needed some form of supervision when reducing the dimensionality. So we added partial least squares (PLS), regression methods such as ridge and lasso, or supervised PCAs approaches. Still, I felt that we could capture complex features of news data better with more expressive statistical approaches, which led directly to the autoencoder branch of neural networks.

Several software solutions in this space are basically at your fingertips these days, which means you can implement a neural network in no time. Tech-wise, I employed PyTorch, a popular open-source deep learning package. In this framework I coded up a custom architecture for autoencoders with supervisions (supervised autoencoders) that integrates both dimensionality reduction and supervised learning. In simple words, this particular architecture allows the model to learn compact sentiment data representations that are already well adapted to our financial use case by way of the supervisions.

What were your findings?

On the positive side, training and coding of the neural networks was much easier than originally thought. Above all, the analysis highlighted the added value of customisation in statistical signal construction. The implementation of custom loss functions in particular resulted in data representations that we liked much better than those of off-the-shelf approaches. Across the board, the results in different use cases robustly favoured the neural networks. Interestingly, the complexity of the neural networks in terms of the depth was of second order. Also, while it was hard to get something significantly better than standard approaches after accounting for multiple selection issues etc., the neural networks produced data representations that were more relevant to us practitioners.

How do you apply these insights at Quoniam in the investment process?

While we are looking at including neural network derived data representations into investment strategies, the proposed approach will be at first most helpful in the research process itself. One example is the review of datasets containing text-derived numerical sentiment scores for many topics. The framework outlined in the paper gives us a very competitive benchmark approach that can directly produce actionable signals.

Explore the paper:

Learning deep news sentiment representations for macro-finance