Navigating the vast amount of information in corporate filings can be overwhelming due to their complexity and sheer volume. Computext simplifies this process by analyzing filings from both domestic and international companies, identifying and tagging sentences relevant to specific financial topics. Leveraging data from Context Analytics' Machine Readable Filings (MRF), which includes over 1.5 million documents, the Global Machine Readable Filings (GMRF) with over 1.25 million documents, and S&P Earnings Call Transcripts, Computext focuses on extracting key financial insights. These insights cover items from Income Statements, Balance Sheets, and Cash Flow Statements, such as Sales, Inventory, Operating Cash Flow, etc. Initially Computext focused on around 20 items, but now covers 38 and continues to expand. Each sentence is labeled with specific financial items, relevant context, sentiment indicators, and identifiers for the document and company.
Overview
We conducted an analysis on the Russell 3000 universe using Computext extractions over a period starting from the beginning of 2014, encompassing over ten years of data. The goal was to assess the long-term impact of the sentiment extracted from these government filings.
Methodology
Our approach involved using Computext to identify positive and negative phrases or sentences from government filings for each company, aggregated monthly across all 38 items. The sentiment score was calculated using the following formula:
Companies with fewer than 10 extractions in a month were excluded to ensure sufficient data for Score calculation. Each month, companies were grouped into quintiles based on their sentiment scores:
- Quintile 5: Top 20% of scores (most positive sentiment)
- Quintile 1: Bottom 20% of scores (most negative sentiment)
Positions were entered on the last day of each month and held for three months, with adjustments made for new filings within the holding period. This process was repeated monthly over the ten-year period.
Improvement and Results
Initially, our research in April utilized 23 Parent Topics, which yielded promising results. We later added 15 more Parent Topics, increasing the total to 38, to provide more depth and coverage in analyzing these extensive documents.
With the addition of 15 new topics, we observed improvements in our monthly strategy:
- Wider Quintile Spread: The performance difference between the top and bottom quintiles increased.
- Improved Risk-Adjusted Returns: Metrics such as the Sharpe and Sortino ratios for the top quintiles showed enhancement.
- Increased Signal Frequency: More companies met the criteria of having over 10 phrases per month, providing a larger dataset.
- Expanded Coverage: In the Russell 3000, over 2,500 companies consistently had more than 10 phrases extracted per month.
These developments suggest that the additional Parent Topics contributed to better differentiation among sentiment scores.
Conclusion
The research indicates that sentiment extracted from government filings using Computext may serve as a useful indicator of long-term financial performance. The expansion to 38 Parent Topics has added depth to the analysis and improved the results. As Computext continues to evolve, further enhancements in long-term performance prediction might be achievable. For more information, visit www.contextanalytics-ai.com