Graham

Our AI software, which we have nicknamed "Graham" after Benjamin Graham, conducts bottom-up fundamental research on each and every public company, assesses its long-term prospects, and estimates its net present (i.e., intrinsic) value to identify and invest in undervalued securities, eliminating the need for human analysts.

Each point in this visualization of "deep representations" is one public US company evaluated on a reporting date by Graham's deep learning system, in this case spanning a five-year period. Each deep representation is a high-dimensional vector of floating-point numbers, which we map to two dimensions via the Barnes-Hut variant of the t-SNE visualization algorithm. The more similar Graham considers two point-in-time companies, the closer their location in the space of deep representations, and the closer their location in the two-dimensional approximation shown here. Each point is colored by a normalized measure of future business profitability, ranging from dark red (most unprofitable) to dark blue (most profitable).

Made Possible by Deep Learning

Graham's key component is a deep learning system that evaluates public companies with computational models composed of multiple processing layers that learn representations of company data at multiple levels of abstraction. Unlike simpler forms of machine learning that discover relationships between variables, this deep learning system transforms raw company data into computer representations of different kinds of businesses. Businesses that have similar high-level features or characteristics get similar representations, and vice versa. The visualization on this page shows that the representations learned by this system are clustered in a manner useful for making predictions about companies (click on the image to zoom in).

This type of machine learning process is commonly called deep representation learning. It is analogous to the complicated cognitive mechanisms by which a human research analyst gradually transforms company data into abstract mental representations. In the analyst's brain, such representations manifest as patterns of electrochemical neuron activations. In Graham's deep learning system, the representations consist of patterns of neuron activation values in the deeper layers of an artificial neural network, with each activation value indicating the extent to which a high-level feature or characteristic identified by the deep learning system is present in a company.

Whereas a human analyst's mental representations will typically be associated with preexisting labels such as, say, "cyclical business" or "mature company," the computer representations learned by Graham may not necessarily be associated with any preexisting human labels. Indeed, Graham may be identifying high-level business features or characteristics for which there is currently no descriptive terminology.

For a non-technical explanation of deep representation learning, please click here. For a technical introduction to the subject, please see the chapter titled "Representation Learning" of the textbook, "Deep Learning," by Ian Goodfellow, Yoshua Bengio and Aaron Courville, MIT Press, 2016 available for free online. For more information on Graham and our automated investment process, please contact us.