Machines Learn Long Term Fundamental Investing: Part 3
We last left off with Part 2 with our introduction to Joe v1.0 who got a lot smarter and started learning from historical case studies rather than just from his own experience which costs money!
Now, on Jan 1, 2007, Joe takes his 6,000 * 2 predictions (one confidence probability that it’s a good long and one confidence % that it’s a good short net of borrow costs) and builds a portfolio with his $10mm in his brokerage account of his highest conviction 2% longs and the highest conviction 4% shorts of the market. He’s careful to keep his exposure market neutral and not attempt to put a lot of money into illiquid or impossible to borrow stocks. He puts in his trades for the week, trying to minimize market impact by never being more than 5% of the volume: some get executed fully, some partially because the price moves outside his limit order. The next week Jan 8, 2007 he incorporates what he learned the previous week (he’s got a whole week of new case studies now!) and then rebalances his portfolio to be the best 2% longs and best 4% shorts (with a slight buffer to minimize trading costs and we track his returns on “real” money – in the simulated, Matrix-like world. Nearly 11 “years” go by and Joe v1.0 has compounded money at 22% per year with his worst drawdown being about 15% when there was a really sharp rally that hurt his short book especially.
Remember though, Joe v1.0 is pure software and we built The Matrix and rewind or fast-forward time (though Joe has no memory of anything in the future so he isn’t the wiser) and re-run history as many times as we want. How do we measure how good of an investor Joe v1.0 is? Well watching his entire career in the super realistic Matrix is pretty good and about the most you could ever hope to ask from any investment manager. But let’s take advantage of the tools available and go one step further. Let’s say you meet an investment manager and he’s got great returns but you’re skeptical: he’s had a big chunk of his portfolio in Amazon for the last 10 years! That’s been the best performing stock – was he super smart or just lucky? The simulation of the entire investment process is extremely important to do correctly for a strategy that incorporates shorting. As legendary investor Howard Marks reminds us about the “6 foot tall man who drowned in the 5 foot average depth river” – even having a valid 90% confidence in a short does not protect from it increasing 10X along the way and bankrupting you.
Let’s try this: let’s make 100 parallel Matrix simulated worlds and dropout a random 20% of all companies that ever existed in each one. No two are exactly alike. In one world there is no Coca-Cola case study for Joe to learn from, another one doesn’t have an Amazon stock that Joe can buy because it doesn’t exist and he’s never even heard of it. How is Joe’s performance in each of these slightly different parallel simulated worlds? Well if he just got lucky with his performance: you’d expect the outcomes to be wildly different right? If he in fact is a deeply good fundamental investor then the outcomes should be pretty similar.
The opportunity set is a bit different so it will never be precisely identical but if his long term track record is always between 17% CAGR and 19% CAGR in ALL 100 worlds (always lower than 22% because the entire opportunity set is smaller than the “real world” of course) then that’s a pretty good sign of a robust investment process. To be conservative, let’s take the worst track record out of the 100 parallel Joe v1.0’s as our metric for how good Joe v1.0 as an investor. The only tools and information we gave Joe when he was ‘born’ Jan 1, 2007 in each world was telling him what we thought the important data points were that we as human fundamental investors believe to underlie actual cause-effect relationships in investing. So net margins are in, first letter of CEO’s last name is out. Joe had access to all the case studies to learn on his own and let the actual data speak for itself about what works to predict the next investment.
We also could look more deeply at Joe’s specific predictions to see if they really make sense to us based on the cause-effect relationships in business and investing. Here’s an example of a stock Joe picked: 1) higher than average revenue growth, 2) higher than average margins, 3) lower than average P/E multiple, and 4) really high and growing accounts receivable. Looks like a lot to like here – especially to a simple ‘value’ investor but what about the A/R issue? Joe is short this stock and 70% sure it’s a good short. Interesting – but think about it: is growth good or bad if you’re not getting paid by your customers? Pretty insightful. And literally an impossible relationship to mathematically capture with a linear model like those used by traditional ‘quant’ investors.
Joe v1.0 has one main limitation – he only knows how to read the numbers. And while accounting is the language of business, the performance of a business is driven by the people running it. And the stock price moves up and down with the supply/demand dynamics that are external to it. So while Joe may be a highly optimized fundamental investor, there’s more to the world than just the company specific fundamentals.
Joe v2.0 has access to more than just the Library of Congress, he also has all the insider trading track records, corporate transaction records, track records of every executive and director at every other company they’ve been at before, and all the other structured data that’s publicly available for US public companies over the last 20 years. Joe v2.0 also has ‘brothers’ – actually thousands of them that all work as hard as him but specialize in a certain type of stock (i.e. industry or market segment). We invest based on a confidence weighted algorithm that synthesizes their different predictions.
We’ll introduce Joe v3.0 and our v4.0 platform that incorporates the domain expertise from dozens of experienced fundamental investors in our next letter.
For the whole story and more - please click below (or to the right) for Exclusive Content.
