TL;DR
This paper introduces a novel "FedSent Index" to quantify the sentiment expressed in Federal Reserve (FED) communications (minutes and speeches). They demonstrate that this sentiment index significantly predicts major currency trends, particularly for USD/EUR, and can even Granger-cause daily exchange rate changes, offering valuable insights for currency market investors.
What Problem Does This Paper Address?
In the complex world of financial markets, understanding the drivers of currency fluctuations is paramount for investors. While central bank communications, especially from the Federal Reserve, are known to influence markets, extracting actionable insights from these lengthy documents has been challenging. Previous research often focused on broad sentiment analysis of entire documents or specific topics, without directly linking it to individual currency movements or assessing short-term Forex market implications. This paper aims to bridge this gap by developing a method to quantify FED sentiment and explicitly test its predictive power on major currency exchange rates.
Methods
The core of this research lies in developing the FedSent Index (FSI). The authors leverage a pre-trained language model, FinBERT, to classify individual sentences within FED minutes and speeches as positive, negative, or neutral. For each document and for specific currencies (EUR, USD, GBP, JPY, CNY), they aggregate these sentence-level sentiments into a document-level sentiment score.
Two aggregation methods are used: mean and median. To account for how specific a document is to a particular currency, they also compute a specificity coefficient ($\beta_{ctd}$), which measures the proportion of sentences mentioning a currency relative to the total sentences in the document. The final FedSent Index is a smoothed version of the aggregated sentiment, multiplied by this specificity coefficient.
To handle the irregular timing of FED communications and the daily nature of Forex data, several imputation methods are employed to create daily sentiment series:
- Ffill: Forward-filling the last available sentiment score.
- Exponential Decay: Gradually decreasing the sentiment's influence over time.
- Most Recent Value: Using the last available sentiment.
- Delta Median: Using the median of previous sentiment values associated with similar currency changes.
Finally, they employ Pearson correlation to assess the relationship between the sentiment index and daily currency changes (with various lags) and conduct Granger causality tests to establish whether sentiment "causes" currency movements.
Datasets
The study utilizes two primary datasets:
- Federal Reserve Communications: This dataset comprises speeches, statements, and minutes from the Federal Reserve, collected from its official website. The data spans from 1993 to September 2023, yielding 1,671 speeches, 252 minutes, and 224 statements.
- Forex Data: The USD/EUR exchange rate time series was obtained from the official site of Banca di Italia, covering the period from January 1999 to September 2023.
The authors note that statements were disregarded due to their limited number of FX mentions.
Key Insights and Ideas
- FED Minutes Matter Most: While both minutes and speeches contain sentiment, the minutes appear to have a stronger and more consistent influence on currency markets, likely due to their more structured and comprehensive nature.
- Sentiment Predicts Trends: The FedSent Index, particularly when smoothed and considering specificity, shows a significant positive correlation with future currency movements, supporting the hypothesis that positive sentiment leads to currency appreciation and vice-versa.
- Lagged Impact: The influence of FED sentiment isn't always immediate. The study reveals that sentiment can Granger-cause currency changes with lags of up to four days, suggesting that investors should consider a window of opportunity after communications are released.
- USD/EUR Focus: The EUR and USD are the most frequently mentioned currencies in FED communications, making their sentiment scores more robust and continuous, thus a primary focus for the analysis.
- Specificity is Key: Simply having a sentiment score isn't enough; understanding how directly the communication addresses a specific currency (via the specificity coefficient) enhances its predictive power.
Key Results
- Significant Correlations: The study found positive correlations between FED minutes sentiment and USD currency changes, with the highest correlation (26.45%) observed for the
FSIµ(c,t,d) × ẞctdscore using the Delta Median imputation method. - Granger Causality Established: For the USD/EUR exchange rate, the sentiment index from FED minutes was found to Granger-cause daily percentage changes, particularly with lags of 1 to 4 days. For instance, using the Delta Median imputation, p-values below the 0.05 threshold were consistently observed for lags 1-4, indicating a statistically significant causal relationship.
- Imputation Method Matters: Different imputation methods yielded varying correlation strengths, with "Delta Median" and "Most Recent Value" often showing higher correlations, suggesting that how missing sentiment data is handled significantly impacts results.
- Speeches Less Impactful: While speeches also contain sentiment, their correlation with currency movements was generally weaker, suggesting they are less specific or their impact is diluted compared to minutes.
Limitations and Future Work
The authors acknowledge several limitations:
- Focus on USD/EUR: The analysis primarily focused on the USD/EUR pair due to data availability and frequency. Future work could explore other currency pairs.
- Exogenous Factors: The study did not incorporate other crucial factors that influence currency markets, such as GDP, inflation rates, unemployment levels, or communications from other central banks.
- NLP Model Limitations: While FinBERT is powerful, sentiment analysis can still be nuanced, and the model might not capture all subtleties of human language.
Future research directions suggested include:
- Expanding the analysis to include other major currencies and central banks.
- Integrating additional economic indicators and geopolitical events into the sentiment analysis framework.
- Exploring more advanced NLP techniques for sentiment extraction and topic modeling.
- Investigating the impact of different communication channels (e.g., press conferences, social media) on currency markets.
Summary updated