We approach investing with a value-oriented process. That is, any investment must consider that stocks represent equity ownership shares in a business. Those businesses are definitionally fairly valued as the net present value of all future free cash flows. In the short-term, the market tends to overreact to exogenous noise (product announcements, news, earnings reports, etc.), disconnecting a stock’s price from the fair value of that corresponding business. Opportunity, therefore, lies in delineating the difference between price and value. We believe with proper valuation techniques and consistent discipline, we can take advantage of short-term mispricings because the market will almost always converge on fair value over the long-term…at some point…eventually.

In the investing world, machine learning and “quants” have typically been utilized for identifying arbitrage opportunities around this exogenous noise to execute many short-term trades on high leverage. Machine learning excels at identifying patterns when given enough data. It just so happens that relatively short-term trading allows you to use a ton of data, because there are simply more “ticks” from which you can sample. However, short-term trading also often requires enough liquidity in the markets to reduce slippage and keep trades profitable. Recently, with a surge of data and liquidity, these ML/quant strategies have also exploded in number. The vast number of these strategies do not consider “value” in any traditional (read: Graham and Dodd) sense. In fact, many of them pile into a business when its price has increased over the past weeks or months! This deluge of short-term focused quants may create an even wider disconnect of prices from value.

It seems that very few firms have successfully used machine learning to identify true, long-term value opportunities. We believe this is a direct result of a lack of the two qualities (liquidity and data) that have made the quant surge possible for short-term trading. Long-term value investments often benefit from small capitalization, spinoffs, restructurings, and other seemingly precarious situations. These cases characteristically come with a dearth of liquidity and data, so if we want to apply machine learning to value investing, the process must look quite different than what has emerged as of late in the quant world. Moreover, an investor must do something differently than everyone else in order to outperform, and that’s exactly what we seek to do.

Machine learning, we’d like to introduce you to long-term value investing. The machine learning problem we must frame and solve is: how can we, on average and over time, identify good businesses that will compound and appreciate over a relatively long-term duration? Kind of a mouthful and it’s still probably easier said than done. So, to solve the problem, we must throw a bunch of buzzwords like natural language processing and crowdsourcing at the problem.

Specifically, we are crowdsourcing write-ups and pitches of long-term, global equity investment ideas to first generate a repository of unstructured language data. This repository of real peoples’ submissions acts as an initial high-quality filter of ideas because we believe humans still have the edge in synthesizing and quantifying unavailable data that pure quants inherently can’t capitalize on (or at least not yet). Then, we can use natural language processing in order to decompose the valuation process for any idea, identifying why the proposed investment opportunity exists (e.g. spinoff, restructuring, negative press, etc.). With typical market and fundamentals data, we also have complementary models that allow us to consider sector, recent income statement changes, cash flow profiles, and many other fields that are historically consistent with long-term outperformers.

We can’t simply search for mass correlations over a gigantic dataset; instead, we need to encode the “rules” and heuristics of value investing into a set of algorithms that all interact with each other to arrive at a final, reproducible investment decision. We like to think of this as deploying a systematic layer on top of qualitative rigor (i.e. machine learning models trained on a sample of investment community submissions). Through this process, black-box machine learning models become powerful and interpretable tools to develop theories around true, long-term value investing.