Deception And Truth Analysis (D.A.T.A.) is a proprietary library of scientifically-based algorithms that rapidly scour financial texts for deception.
Deception And Truth Analysis - D.A.T.A.
HOW D.A.T.A. WORKS
Language Fingerprints Examined. Our algorithms work by examining how language is used, and not what language is used. Deception And Truth Analysis (D.A.T.A.) scrutinizes over 45 unique language fingerprints, making it exceedingly difficult to spoof its algorithms. Other text-based deception detectors look for the presence of "red flag" words. Sadly, recent research (2020) shows that the advantage from searching texts for these words has created an arms race, with companies now hiring speech and language coaches in order to excise offending "red flag" language from their texts and speech. Deception And Truth Analysis results are re-scaled with most falling between -100 (very likely deceptive) to +100 (very likely truthful).
Algorithms Developed Out-of-Sample. With the exception of one algorithm, D.A.T.A. and its algorithms were developed out-of-sample. That is evaluating how people deceive using language, in general, rather than how they use language to deceive in finance and investing. These general deception detection algorithms are then applied to financial communications to render a deceptiveness or truthfulness score. It is our feeling that using a combination of algorithms that look at different language fingerprints significantly improves the accuracy and is demonstrated by our research.
Scientifically Demonstrated Accuracy. Each algorithm in the library is supported by scientific research and its ability to surface deception is statistically significant and in excess of the capabilities of people. One of the algorithms has been measured with an accuracy of 72.74% vs. a January 2021 study that placed human capabilities of detecting deception in texts at an even 50.0% chance level. Additionally in experiments D.A.T.A. is able to identify all 10 of the largest scandals in global corporate history and with an average 6.6 years lead time. Last in a test of Deception And Truth Analysis when used for security selection resulted in an average annual outperformance of the equal-weighted DJIA of 104 bps, outperformance in 9 of 11 years, and compounded outperformance of 11.85%. See our Research section for more information.
No overfitting. Look, as investment managers ourselves we are dubious of back-testing and data-mining. We have literally seen and heard it all. For Deception And Truth Analysis we curated scientifically-tested algorithms developed for general deception detection, then applied them to financial data. Using this algorithm library we then looked for those with little overlap with one another in terms of the language fingerprints looked at. That is it. We have run optimizations on D.A.T.A. and in the case of the DJIA research we conducted, it is possible to overfit and optimize to generate a maximum outperformance of 162 bps, outperformance in 9 of 11 years, and compounded outperformance of 19.20%. Instead, the results we report and that we want you to focus on are those generated in deference to Occam's Razor.
Huge Normative Database. A given financial text selected and uploaded by you is compared to a large financial communications database of millions of words. Using this normative data helps us to refine the results and ensures that a financial communication is being compared to others of its kind and not to the diaries of teenagers, for example.
A Separate Fraud Score. A separate fraud score is also provided by comparing the language fingerprints of a communication to those of known fraudulent companies. Because of the chance of extreme capital loss due to fraud, these algorithms are designed to provide more false positives than negatives. These scores also generally range between -100 and +100.
Extreme Scores Tracked. Additional detail is provided about the number of extreme scores to ensure that the result is not overly skewed by one or two language fingerprints.