Fine-tuning Language Models for Equity Trading using RavenPack Annotations

October 24, 2023

We fine-tune FinBERT with RavenPack Annotations to predict stock market movements, delivering an average Information Ratio increase of 48%.

The broader availability of large language models has renewed investors’ interest in sentiment-driven trading strategies. While open-source models adapt well to the task, they rely heavily on quality training data.

In this paper, we report on the impact of refining FinBERT, an open-source language model already fine-tuned on a financial dataset, with the RavenPack Annotations dataset to forecast stock market movements.

Improved Sentiment Analysis and Performance

RavenPack Annotations contains sentiment information on the entities detected in each sentence. Fine-tuning the FinBERT model with this dataset generates a substantial performance boost. When applied to backtesting six investment strategies across global markets, the fine-tuned FinBERT model increases the information ratio by an average of 48% and the annualized returns by 47%.

The results prove robust across all our universes over multiple holding periods, from daily to bi-weekly, with the fine-tuned model consistently outperforming the open-source version.

Watch the video summary

Portfolio IR AR Portfolio Size
Model Raw Fine-tuned Raw Fine-tuned Raw Fine-tuned
US Mid/Large-Cap 0.80 1.27 4.35% 6.74% 149.5 169.9
US Small-Caps 2.63 3.42 25.55% 29.63% 128.2 160.6
EU Mid/Large-Cap 1.42 2.36 8.82% 14.76% 62.6 65.8
EU Small-Caps 0.99 2.03 9.79% 20.15% 15.8 18.4
APAC Mid/Large-Cap 2.86 3.07 19.05% 23.29% 47.4 49.0
APAC Small-Caps 2.47 2.94 23.91% 28.12% 26.4 30.5

Download the white paper to explore the methodology and review applicable use cases.




By submitting this form, you agree to RavenPack's terms of service and privacy policy.

Data Insights

Read More