A Dynamic View of Intrinsic Value
Meson Gravity Approach: A Dynamic View of Intrinsic Value
Our investment strategy has always been to buy companies at a big discount to their intrinsic value and either a) be patient or b) act as a catalyst to close the gap or increase the intrinsic value over time. Sounds simple – but how do you determine intrinsic value? Even if you could determine intrinsic value today precisely, what if the company changes? The static view of intrinsic value is so incomplete that it can be a dangerous concept. The world is dynamic and we can only see a little bit up the road – even companies themselves can’t predict their own revenues a year out with the benefit of total inside information. We can estimate a range for intrinsic values but that too is limited when framed in linear thinking tied to today.
A 30-year lease on a commercial building has a pretty clear intrinsic value but the range for operating businesses is as wide as ever. The core value of businesses is increasingly the intangible (i.e. informational) component and decreasingly the stack of bricks or factory floor with easy to observe GAAP accounting metrics. The first derivative of value is the quality of the people and the thousands of decisions that get made each day to cumulatively determine future value.
We aim to compete with other human investors while utilizing our machine learning tools to achieve superior consistency and depth of research for our own fundamentally driven investment process. As long as there are emotional and biased people in the market and Jim Cramer, et al. have the attention of investors wielding meaningful capital and 95% of sell-side broker recommendations are “BUY” etc. there will be an opportunity set for a cool rational approach. Stocks are not bought, they are sold.
Core tenets of our machine learning enabled strategy include:
1) A long term fundamental approach over a year or longer, not weeks or months with short term traders. This timescale is uncompetitive with quant investors as few have the long term conviction about a business to ride out a month or quarter that isn’t ‘working’ in the stock.
2) We focus on the deeper causal factors in our data – people, business quality over time, and supply/demand dynamics in an industry. What drives the change in intrinsic value over time for a company? Valuation metrics are important but generally too obvious to gain an edge.
3) We short for the long term: economic gravity always wins in the end with low quality businesses run by self-motivated people.
Requirements to follow these tenets and the barriers to entry are:
1) A long term approach requires non-linear models to make investment decisions – i.e. machine learning, not merely statistics. The technology to implement this has only recently become feasible at a reasonable cost and the software engineering requirements are substantial. As an investment strategy – this type of modeling can capture nuances about a company the same way that a human investor does. This is in stark contrast to ‘smart beta’ or ‘factor’ investing where the characteristics are easily measured (such as low price/book value) and arbitraged away .
2) Knowing what data to look at requires real domain expertise as a long term investor and the ability to translate that into the same language that a machine can understand: typically orthogonal skillsets. This is the most proprietary piece of our investment process and requires meaningful effort to gather and structure data that is not available from commercial vendors. We agree with Google’s Alon Halevy that the hardest problems and biggest breakthroughs come from integrating different datasets into one multifaceted view of reality.
3) To maintain long term conviction in shorts, diversification is required so a short going against you doesn’t need to be reacted against adversely solely due to price action. A problem still remains: if you are right on 99% of your shorts but 1% of time the time you short Amazon in 2003, you lose. To validate our models, we had to build a proprietary simulation architecture to realistically play out thousands of alternate versions of history. The computing infrastructure required for this would have cost millions of dollars before recent cost reductions and software advances in cloud computing. As a quick aside – why short at all if it’s so hard? A) Despite the rising index, the performance is mostly from the big winners and most stocks perform worse than T-bills . B) We are trying to predict the future using historical data – great businesses are almost always doing something new and harder to predict whereas failure is much easier to predict.
Note that this approach is very different from how typical quant funds work: rather than running 1000 back-tests to search for what strategy would have worked, we forward-test our fundamental ideas with many variations on history to reduce tail risks before anything starts. Risk control is paramount and simple metrics like VaR are inadequate. We will undoubtedly encounter new market situations that have never happened before and will respond our best, but at least we can start with knowing we won’t likely repeat the same mistakes from history.
A Challenging Market Environment for Long-Only Stock Pickers, Better than Ever for Entrepreneurs:
The current environment of high valuations and rapid technology change should provide unprecedented tailwinds to entrepreneurs who can harness them. Cost of growth capital is as low as any point in history and the amplification effects of technology make human willpower and intelligence more economically potent than ever. On the short side – increased competition and technological disruption is making the lifecycle of poorly run companies shorter than ever. The inflated valuations across the market allow for attractive entry points, decreasing upside risk for these companies going forward.
This is part 3 of 3 of an article about how machine learning is impacting the investing world. Click below to learn more.
