This repository holds the deception model built in support of an academic research project and resulting manuscript that explored the effects of CEO deception on analyst ratings following a cognitive load inducing situation (in this case, largely quarterly earnings calls).

Please consider citing the original work in which this model was first developed and used:

Hyde S., Bachura E., Bundy J., Greatz R., Sanders WM G., (forthcoming). The tangled webs we weave: Examining the effects of CEO deception on analyst recommendations. The Strategic Management Journal

Two versions of the model can be found in the models directory. One of these is a scikit-learn version of the model that has been serialized using joblib. The other version is a pytorch equivalent model that can output identical classification labels.

I have provided two sample python scripts demonstrating how to load these models and running a test against already known scores to ensure you are producing the correct values. Please note that it will be necessary to load the appropriate libraries prior to running the script. This can be accomplished by installing them with the requirements.txt file: e.g. - pip install -r requirements.txt (I recommend doing this in a virtual environment for python, such as venv).

These models are available under the ASL license that provisions use for non-commercial academic research (see license for more information).

I recommend reading the original paper referenced above to fully understand how this model was developed. However, in lieu of this, it is important to understand that usage of this model carries with it some assumptions.

1. • Sufficient semantic content (model versions were all developed with a minimum of 150 words of content)
2. • Semantic content expressed in a cognitive load inducing situation. The spoken text is contextualized by a situation in which the spoken text was generated while the speaker was necessarily using cognitive functions for that generation, e.g. - recalling events, processing new information, formulating responses dynamically.
3. • Any potential deception present in the semantic content is the result of non-trivial concerns, e.g. - there are consequences to the discovery of any potential deception that undesirable for the potential deceiver.
4. • Linguistic Inquiry Word Count (LIWC) values from the LIWC 2015 dictionary. These are the input feature space values originally used to develop these models. Newer and more complex models have been developed with greater performance and without the licensing limitation of having to use LIWC features: See next section for more.

Commercialized models of a non-derivative nature (created separately and functionally dissimilar) that can use semantic, phonetic, and mixed-modality feature spaces have been developed by the creator of these models. These commercially available models do not require LIWC input features, can handle shorter responses, and have been validated in a greater range of contexts. Please contact [email protected] directly for more information.