Extract the Beliefs That Drive Markets
Illustrative comparison. The same passage, three classification approaches.
01 / Definition
Financial text contains both facts and judgment. Revenue can rise while an author remains bearish. Weak earnings can accompany a bullish outlook. The facts and the opinion are distinct pieces of information.
Sentiment measures the author’s subjective attitude toward a specific asset, expressed through opinion or emotion. It answers a direct question: What view does the author express about this asset?
Investors encounter these views in commentary, analyst notes and social discussion. The measure records the target asset and the direction of the author’s expressed view.
The object
Opinions and emotions expressed by the author. Favorable or unfavorable facts alone receive a neutral label.
The target
Each label applies to the target asset. A bullish view of one company does not extend to every company mentioned in the text.
The value
The measure separates stated optimism or pessimism from factual reporting. It supports analysis of belief formation, disagreement and market participation.
Definition and decision rules follow the working paper, Separating Sentiment from Fundamentals Using a Large Language Model: How Does Media Sentiment Speak to the Crowd?
02 / Measurement
The three labels record the direction of sentiment, not its strength.
Negative
−1An unfavorable opinion or negative emotion toward the target asset.
Neutral
0No clear subjective direction toward the target asset. This category includes factual reporting and ambiguous expression.
Positive
+1A favorable opinion or positive emotion toward the target asset.
Text × Target asset
Interpret the author’s opinion first. Use emotion when the opinion is neutral. Apply the same rules to short posts and long commentary.
Asset × Publication date
Aggregate positive, negative and neutral labels by asset and publication date. Report category shares, text counts and changes over time. State the denominator.
Each text receives −1, 0 or +1. Averages and category shares summarize these labels across texts. When short- and long-horizon views conflict, the annotation rule prioritizes the short-horizon view. If the short-horizon view is neutral, the longer-horizon view applies.
Five annotators trained in finance define the decision rules. Task-specific fine-tuning trains the model to apply them. Independent evaluation tests performance on ordinary and difficult texts.
Difficult posts make up half of both the training corpus and the benchmark. The benchmark is separate from training and is evaluated after the model and prompt are fixed. Accuracy on external evaluation samples appears below.
| Evaluation set | Texts | Accuracy |
|---|---|---|
| English translation of benchmark | 600 | 86.83% |
| StockTwits | 210 | 90.00% |
| Long-form financial texts | 135 | 90.37% |
Source: working paper, Methodology and Validating the Sentiment Classification. Accuracy is measured on the specified evaluation samples. It does not estimate accuracy across the full production archive.
03 / Descriptive statistics
Explore the professional commentary sample and the retail discussion archive. Each sample has its own source, time span and classification denominator.
Publication year · Share of classified texts
Professional commentary spans 2 January 1992 to 29 November 2024. Annual shares pool article classifications by publication year. The working paper’s pipeline assigns a neutral label to 324 articles with no valid raw inference label.
Same articles · Two measurement targets
The two measures agree on 41.2% of classifications. Negative dictionary tone often accompanies a neutral or positive author view. The matrix shows how their labels differ.
6,018 articles combine negative dictionary tone with positive author sentiment.
Rows classify dictionary tone at threshold 0.25. Columns classify asset-specific author sentiment. Each cell percentage uses its row total. The matrix compares the two measures; it does not test classification accuracy.
Professional commentary · Source composition
Each bar shows sentiment shares for one publication and is normalized to the same length. Article counts indicate source coverage.
Professional commentary. The sample contains 74,525 articles covering 228 companies, from 2 January 1992 to 29 November 2024. It includes firms that entered the largest 100 U.S. companies at least once during the collection period. Article-level descriptive statistics use the working paper’s raw-article inference pipeline, which assigns a neutral label to 324 articles with missing labels.
Retail discussion. The archive contains 365,701,929 dated posts in 2010–2025. Of these, 358,369,728 have valid sentiment labels. Annual shares exclude 7,332,201 posts with no valid label. The 2025 sample ends on 7 August. Publication dates include non-trading days. Statistics are grouped by publication year.
Source identity. Retail statistics use the Eastmoney Guba discussion archive. Commentary statistics use the working paper’s article sample. Counts are calculated separately for the two samples.
Download descriptive statistics (JSON)| Year | Classified texts | Negative | Neutral | Positive |
|---|---|---|---|---|
| 1992 | 1,593 | 319 | 695 | 579 |
| 1993 | 1,814 | 382 | 732 | 700 |
| 1994 | 4,174 | 447 | 2,698 | 1,029 |
| 1995 | 5,248 | 818 | 2,866 | 1,564 |
| 1996 | 5,731 | 1,072 | 2,845 | 1,814 |
| 1997 | 6,231 | 1,096 | 3,405 | 1,730 |
| 1998 | 6,104 | 1,219 | 3,339 | 1,546 |
| 1999 | 6,055 | 1,034 | 3,435 | 1,586 |
| 2000 | 1,018 | 218 | 596 | 204 |
| 2001 | 1,713 | 451 | 878 | 384 |
| 2002 | 3,011 | 793 | 1,487 | 731 |
| 2003 | 1,728 | 427 | 877 | 424 |
| 2004 | 1,047 | 239 | 572 | 236 |
| 2005 | 1,460 | 292 | 807 | 361 |
| 2006 | 1,324 | 270 | 729 | 325 |
| 2007 | 1,671 | 372 | 912 | 387 |
| 2008 | 2,239 | 693 | 1,114 | 432 |
| 2009 | 2,668 | 665 | 1,320 | 683 |
| 2010 | 2,367 | 505 | 1,266 | 596 |
| 2011 | 1,935 | 448 | 1,051 | 436 |
| 2012 | 1,777 | 426 | 977 | 374 |
| 2013 | 1,379 | 320 | 781 | 278 |
| 2014 | 1,672 | 356 | 1,035 | 281 |
| 2015 | 1,430 | 330 | 822 | 278 |
| 2016 | 1,252 | 342 | 676 | 234 |
| 2017 | 1,478 | 329 | 850 | 299 |
| 2018 | 1,418 | 332 | 823 | 263 |
| 2019 | 1,095 | 352 | 558 | 185 |
| 2020 | 674 | 169 | 327 | 178 |
| 2021 | 760 | 205 | 388 | 167 |
| 2022 | 877 | 206 | 488 | 183 |
| 2023 | 804 | 213 | 386 | 205 |
| 2024 | 778 | 247 | 335 | 196 |
| Year | Classified texts | Negative | Neutral | Positive |
|---|---|---|---|---|
| 2010 | 204,469 | 21,572 | 147,714 | 35,183 |
| 2011 | 307,364 | 41,413 | 209,416 | 56,535 |
| 2012 | 195,806 | 45,306 | 91,165 | 59,335 |
| 2013 | 1,222,826 | 151,499 | 887,275 | 184,052 |
| 2014 | 2,505,559 | 607,057 | 1,319,875 | 578,627 |
| 2015 | 6,570,314 | 1,925,860 | 2,970,860 | 1,673,594 |
| 2016 | 16,767,574 | 5,523,164 | 7,242,726 | 4,001,684 |
| 2017 | 25,101,583 | 7,823,695 | 12,080,444 | 5,197,444 |
| 2018 | 23,174,433 | 8,102,147 | 10,251,651 | 4,820,635 |
| 2019 | 31,720,576 | 11,171,244 | 13,107,445 | 7,441,887 |
| 2020 | 47,670,467 | 17,265,522 | 19,186,948 | 11,217,997 |
| 2021 | 55,865,998 | 20,079,217 | 22,965,865 | 12,820,916 |
| 2022 | 48,241,198 | 18,011,895 | 19,975,940 | 10,253,363 |
| 2023 | 36,501,573 | 12,019,035 | 17,017,927 | 7,464,611 |
| 2024 | 36,177,807 | 11,752,370 | 16,858,004 | 7,567,433 |
| 2025 | 26,142,181 | 8,481,325 | 12,124,093 | 5,536,763 |