Deux ex machina

An extract from our 2016 Annual Letter

*admittedly Man designs the Machine so ultimately he wins; as per the English claim that they always win rugby as they invented it.

Part I: Man versus Machine - Man 0, Machine 1*

One thing which strikes us, apart from the whoosh as the years fly by, is how revolutionary recent advances in computing, data science and machine learning are.

Consider the increasing number of ‘high order’ pattern recognition tasks where computers outperform humans: chess, poker, quiz shows, cancer diagnosis, driving, et cetera.

“It is a renaissance, it is a golden age. We are now solving problems with machine learning and AI that were … in the realm of science fiction.”

-          Jeff Bezos, December 2016

Even though we live and breathe this (exponential) world, we failed to foresee the pace of change. But we are not complaining – these developments are all ideal for G2, as we continue to apply the latest in AI/machine learning to stock picking.

A personal anecdote to give a sense of history: when one of the G2 partners was working on his PhD (just) ten years ago, experts were still training computer systems. Now, computer systems are training the experts.

Nevertheless we are far from being techno‐utopians; our mindset is much more engineering than evangelical.

In fact 5 years in the business of systematizing stock picking has taught us how far away we are from easy solutions. Like the curmudgeonly old‐timer, we find ourselves increasingly skeptical of expectations and claims made about AI.

Indeed, outside of dating, sexiness is the enemy of results.

It’s getting to the point where we need to preface everything we say with: we don’t mine twitter!

G2 has had no changes in our human (analogue) team, but our investment team is still growing: we have 26 virtual analysts and 45 virtual traders. They are hyper‐competitive but without ego.


Part II: “I’d be a bum on the street with a tin cup if markets were always efficient”

So said Warren Buffett, famously.

Indeed, successful investing requires taking action only when a great business becomes temporarily undervalued. The market is ‘mostly efficient’ rather than ‘always efficient’: the distinction makes all the difference in the world.

Thus, the task of an investor is to constantly search for mispriced assets and pounce when Mr Market becomes irrational.

If this sounds simple, why isn’t it easy? Why do most individual investors and active managers fail to outperform market indices?

The bottom line is that the stock market is, in the jargon, a wicked learning environment. The rules of the game (signals) are hidden in randomness, noise and further obscured by behavioural/institutional biases and emotional responses.

Moreover, great investments are ephemeral, requiring a combination of super‐human patience married with aggressive opportunism when a ‘fat pitch’ presents itself.

No wonder successful stock picking is a rare skill!

Indeed, the average stock picks by professional managers perform worse than the market (research by Inalytics cited in our earlier letter revealed that the average hit ratio is 49.6%, meaning that fund managers, on average, are wrong more than half the time).

While there are many ways to investment heaven – and many investors have demonstrative skill – G2’s approach is to use machines to learn the rules of the game.

That is, we apply machine learning to the task of finding reliable patterns in company data. This goes beyond typical ‘Quant’ approaches based on blunt factor exposures and regression, and instead looks to self‐identify granular patterns in fundamentals at the company level.

In essence we are trying to combine the breadth quant strategies benefit from (1000’s of companies analysed daily) with the depth normally reserved for human analysis.

G2’s system has access to 10,000 data points per company and uses 252 composite variables to elicit reliable patterns. The variables (we call them “features”) span fundamentals, sentiment, analyst expectations, governance and technicals.

The proof is in the pudding: G2’s hit ratio, based on over 1,000 matched trades since inception, stands at 56.6%.

As such, G2’s US Alpha strategy has delivered top decile returns since inception over 4 years ago (data source: Morningstar).

Key to our systematic approach is the layering of independent and uncorrelated models:

1. Virtual Analysts ‐ an ecology of 26 diverse VAs compete (and sometimes co‐operate) to select stocks. Focused on company fundamentals, these algorithms exploit small pockets of mispricing rather than making concentrated factor bets.

2. Virtual Traders ‐ take the VAs’ recommendations and block or take off positions where short‐ and long‐term predicted returns are at odds. Each of the 45 VTs focus on a specific variable: e.g. short interest, price action, insider ownership, analyst upgrades.

3. Virtual PMs use individual model performance of the 26 VA’s as well trader recommendations to build the final portfolio. Net exposure is then adjusted using index and macro level variables.

4. A Virtual Risk Manager has the final say on the overall risk of the proposed portfolio and adjusts gross exposure by specifically assessing tail‐risk (C‐VaR) in over 10,000 Monte Carlo simulations.

The resulting model diversity, we argue, reduces the risk of encountering single points of failure.