Trading the news with dynamic feature hierarchy trees

News is not static — it fluctuates in volume and importance. This affects its trading relevance and impact, especially considering that markets are also constantly evolving. Moreover, investor competition using similar insights can reduce the effectiveness of certain signals, including news sentiment.

Our latest white paper introduces a model that leverages Feature Hierarchy Trees to dynamically create and combine signals extracted from RavenPack News Analytics.

Feature Hierarchy Trees break large sets of data into smaller features and organize them hierarchically based on importance. This enables us to look at numerous dimensions in the RavenPack Edge News Analytics such as event taxonomy, relevance and novelty data from various perspectives. This approach enhances transparency, robustness, and flexibility in the portfolio construction process.

Leveraging this approach, the research revealed that:

• Tracking a broader set of events is becoming increasingly important over time. Guidance events are also considered to be important predictors.

• The dynamics of the model allows it to accommodate changes in news content and market conditions over time. For example, it captures a decay in earnings signals during 2019-2021, or a growth in partnerships events after 2017 (for daily horizons).

• The model provides robustness across various parameter configurations, as well as a flexibility to adjust the balance between signal strength and volume by selecting features with varying confidence and relevance thresholds.

• When applied to long-short equity portfolio construction, our model demonstrates a significant improvement in performance compared to benchmark models, with an average relative increase of 20% in Annualized Returns and 10% in Information Ratio. This outperformance is extended into longer holding periods of up to 1 month.