Introduction
Introduction
Today, we are proud to officially announce our newest product: Computext. Computext searches through filings from both domestic and international companies, tagging relevant sentences for specific financial topics. It utilizes documents from CA's Machine-Readable Filings (MRF), with over 1.5 million documents, Global Machine-Readable Filings (GMRF), with over 1.25 million documents, and S&P Earnings Call Transcripts. The system extracts sentences related to Income Statements, Balance Sheets, and Cash Flow Statements, labeling them with specific financial items, sentiment, and document and company identifiers.
Currently, Computext covers over 20 items, such as Sales, Inventory, and Cost of Goods Sold, with plans to expand to all items covered by Compustat. This tagging process provides users with a quick and concise summary of the sentences driving the company's financial performance.
Overview
The purpose of this blog is to analyze the short-term financial impact of the Computext extractions found in these government filings. We conducted an extensive analysis of the Russell 3000 universe using Computext extractions, covering a period starting from the beginning of 2014 and spanning over ten years of data.
Methodology
- Data:
- Universe: Russell 3000 securities.
- Time Frame: 2014 onwards (10+ years).
- Sentiment Scoring:
- For each day a security has a filing, we aggregate the number of extractions by polarity: Positive Sum, Negative Sum, Total Sum
- Range: [-1, 1], indicating sentiment from highly negative to highly positive.
- The distribution of the Score is shown below, with the X-Axis representing the continuous Score created and the Y-Axis representing the number of extractions. It appears roughly normal, with a noticeable spike at 1, indicating only positive extractions for that company on that day. This consistent distribution throughout the data allowed us to establish thresholds for grouping.
- Grouping:
- Top 20%: Extreme positive scores. This group had 25,282 signals in the 10-year history.
- Bottom 20%: Extreme negative scores. This group had 25,000 signals in the 10-year history.
- Middle 60%: Mostly neutral scores. This group had 76,193 signals in the 10-year history.
- Performance Analysis:
- Calculated the Return for each security for the subsequent 10 market days post-event.
- Adjusted returns using SPY ETF as the market baseline.
- Averaged the returns for each group over the 10-year period and converted to basis points (bps) for comparison.
Findings
- Top 20% Scores:
- Initial Jump: Over 40 bps on the first day following the event.
- Sustained Performance: Average excess returns of over 40 bps for the subsequent 8 market days and just below 40 on the 10th day removed.
- Bottom 20% Scores:
- Initial Drop: Over 20 bps on the first day following the event.
- Continued Decline: Reaching near -40 bps by the 10th market day post-event.
- Middle 60% Scores:
- Performance: Hovered around the SPY baseline with negligible deviation.
Conclusion
The Computext sentiment scoring system demonstrates a strong predictive capability for short-term market returns. Securities with extreme positive sentiment scores show significant positive excess returns, while those with extreme negative sentiment scores exhibit substantial negative excess returns. The middle 60% group aligns closely with market performance, highlighting the neutrality of the majority sentiment. For more information on Computext, visit www.contextanalytics-ai.com or click the button below.
