TLDR: This research paper evaluates the practical application of sentiment analysis in financial trading by backtesting sentiment-based strategies on Dow Jones 30 stocks. Using three machine learning models, the study found that all models generated positive returns, with a regression-based model (RoBERTa+Transfs) achieving the highest cumulative return of 50.63% over 28 months, significantly outperforming a Buy&Hold benchmark. The findings suggest that sentiment analysis, especially with continuous sentiment prediction, can effectively enhance investment strategies and generate positive alpha.
Sentiment analysis, a technique widely used to understand public opinion from product reviews to political discourse, is increasingly being explored for its potential impact on financial markets. While much research in this area has focused on classifying sentiment at a granular level, its practical application in real-world trading strategies has often been overlooked.
A recent study, titled Backtesting Sentiment Signals for Trading: Evaluating the Viability of Alpha Generation from Sentiment Analysis, aims to bridge this gap by rigorously evaluating how sentiment-based trading strategies can generate positive returns, known as ‘alpha’. The researchers conducted a comprehensive backtesting analysis, simulating how these strategies would have performed in historical market conditions.
The study utilized sentiment predictions derived from three different machine learning models: FinBERT and DualGCN, which classify sentiment into discrete categories (negative, neutral, positive), and RoBERTa+Transfs, which predicts sentiment as a continuous value ranging from -1.0 to +1.0. These models were applied to a vast collection of news articles related to the Dow Jones 30 stocks, covering a period from January 1, 2020, to April 30, 2022. This period was specifically chosen to include significant market events, such as the major decline and subsequent recovery of the Dow Jones 30 index.
To make trading decisions, the sentiment scores for each asset were aggregated daily, before the market opened, into a scale from 0 (highly negative) to 100 (exceptionally positive). Based on these aggregated scores, the strategy generated buy, neutral, or sell orders. Specifically, a ‘Buy’ order was issued if the sentiment score exceeded 60, a ‘Sell’ order if it fell below 40, and a ‘Neutral’ order for scores in between. The strategy employed an equal-value order approach, meaning a fixed amount of money ($10,000 per order) was invested or divested for each signal, with an initial capital of $300,000 and a commission fee of 0.05% per trade.
The results of the backtesting analysis were compelling. All three sentiment models generated positive cumulative returns over the 28-month period. Notably, the RoBERTa+Transfs model achieved the highest cumulative return of 50.63%, significantly outperforming the benchmark Buy&Hold strategy, which yielded 26.96%. This superior performance of RoBERTa+Transfs is attributed to its ability to predict sentiment as a continuous value, offering a more nuanced and precise understanding of market sentiment compared to models that rely on discrete classifications.
Also Read:
- Unlocking Financial Intelligence in AI: Introducing FEVO
- Unmasking AI: New Research Advances in Detecting and Attributing Machine-Generated Text
Beyond just returns, the RoBERTa+Transfs strategy also demonstrated superior risk-adjusted performance, as indicated by its higher Calmar, Sharpe, and Sortino ratios. While it exhibited slightly higher annual volatility compared to the other sentiment models, its substantial returns justified the additional risk. The study highlights that sentiment analysis, particularly when using models that provide a continuous range of sentiment, holds significant potential for enhancing investment strategies and improving financial decision-making, offering a valuable tool for investors seeking to generate alpha in dynamic markets.
