Machines Learn Long Term Fundamental Investing: Part 2
Last time we left off with "Joe v0.1" our maybe human maybe machine (or maybe a machine designed to think he's human like Blade Runner...) investor. Continuing the question of how does an investor learn:
Joe v0.2
Joe v0.2 is a bit wiser. He realizes that Joe v0.1 was onto something by looking at more stocks at a time but is less patient with this waiting idea. So Joe v0.2 goes to the Library of Congress which has all the historical public SEC filings of companies that have existed over time – even if they don’t exist today because they were bought out or went bankrupt. After waking up with zero knowledge of investing, he starts in the 1996 section (when the SEC went electronic with EDGAR) and starts reading 10-K and 10-Q filings for every stock in the US. He’s very careful to note the date on it when it was published & publicly available (Dec 31 reports are usually published by March 15 the following year for example). Then after reading all the numbers in each filing, Joe looks up the historical stock price on the day after the published filing date, the valuation metrics (anything that incorporates the stock price: P/E, P/Book, EV/EBITDA, etc.) and for good measure the stock chart prior to that date, while taking extreme care to not peek into the future after that date! Next, based on the data Joe now knows about each company on that date, he makes a prediction for each stock – will it go up, down or no idea over the next year?
Then, Joe v0.2 turns over the page in the stock market listing for the answer! Where was he correct? Where was Joe wrong? He repeats this deliberate practice, case study method while carefully not biasing himself by viewing any future information and works his way from 1996 up to Oct 23, 2016 (i.e. 1 year ago today since the “correct answer” to the case study is 1 year in the future) and only then does he actually build a portfolio on real money. For those counting: that’s about 7,500 companies (there used to be 10,000 public companies in the US, now only 5,000) * 21 years * 4 filings per year = 588,000 case studies just using quarterly numbers. You could even snapshot the investing world every trading day or more frequently since price changes influence valuation metrics every instant. Each case has 50+ important metrics (revenue, margins, growth, return on assets, valuation, momentum, etc.) just on the GAAP financials and we haven’t even got to the conference calls, industry specific metrics, insider trading form 4’s, analyst coverage, executive and board track records, etc.
Joe v0.2 comes up with a prediction for every case study and importantly – he writes down his confidence in that prediction. Really screaming shorts he’s 90% certain the stocks will go down over the next year, other stocks that seem unclear he puts down 51/49 because he can’t form a strong opinion. As you may have guessed: Joe v0.2 does not have time for a girlfriend, or any friends, or a body. Joe v0.2 is literally a machine whose sole purpose is to learn how to invest.
But now back to #2: how do we know that Joe is learning the right things? How do we know that Joe is becoming a better fundamental investor over time? Unlike short term daily trading strategies – we need to wait at least a year to know if we’re “correct” about any particular fundamental long or short pick. Good thing Joe v0.2 keeps track of his confidence levels for each prediction he ever makes…
Joe v1.0
Now we get to Joe v1.0, a full version upgrade! Joe v1.0 is just like v0.2 but he’s not just a machine but now he’s completely made of software. Because he’s just made of software, we can do a couple things: 1) We can make exact copies of Joe and 2) We can simulate Joe’s ‘world’ and he can’t even tell he’s in the Matrix because it’s 100% realistic. Joe v1.0 wakes up one day and its January 1, 2007. He starts just like Joe v0.2 by going down to the library of congress and reading through all the financial metrics, valuation metrics, etc. for every single US public company since 1996 and does his case study learning method on all of them, all 330,000 of them or so (11y * 7500co’s * 4Qs/yr). Then he takes a look at all the companies that trade today, there are fewer than before because of all the PE buyouts and lack of IPOs lately but still over 6,000 stocks to choose from (there were more stocks then than today…). Based on his experience with all the historical case studies he makes his predictions, including his confidence levels. Joe has honed and calibrated those confidence levels after 330k+ case studies to be verifiably accurate in the past – but what if the world changes in the future?
Part 3 coming soon, click below for more information...
