In Favor of Backtest-Driven Investing Strategies

1. When we decide whether or not to buy a stock, or choose what stock to invest in, we pay attention to certain aspects of that stock (low price-to-sales ratio, for example, or strong earnings growth), and we favor stocks with those aspects in part because we’ve learned that such stocks tend to outperform. We are thus informed in our stock choices by past performance data.
2. Because we want to be thus informed, and because we have no data for the future, we must mine the past for as much data as possible. Past data might indeed show, for example, that companies with strong earnings growth and a low price-to-sales ratio tend to do better than other companies. Or it might show the opposite. If, on the other hand, we think past data is useless for getting good future results, it might be better for us not to invest at all. Investing without data is like throwing darts blindfolded: you might still hit the target, but you’re not focused on it, and you might hit something else entirely.
3. The question then becomes how to best use this data. A simulation of a strategy, also known as a backtest, is a way to test the probability of a desired result on the assumption that the data pertaining to the past is more or less the same as the data pertaining to the future. That assumption is extremely problematic, especially since every backtest necessarily involves only a finite set of past data, and since misuse of that data can be psychologically gratifying. But relying on data without backtesting it blinds one to potential pitfalls. For example, past data shows that stocks with low value ratios tend to outperform, but only a backtest will show that following a pure value strategy will result in long periods of underperformance and is subject to far worse drawdowns than a strategy that mixes value with other traits.
4. The components of an investment strategy must be formulated not just from observation of data but from understanding why the data shows what it shows. If there is no coherent explanation based on sound principles of analysis for a conclusion drawn from data, that conclusion should be discarded.
5. The best way to find out whether a backtesting procedure will produce strong results is to do correlation studies. Develop fifty different strategies and see how well their simulated results correlate over many different time periods using different procedures. For example, a five-stock simulation will probably be less correlative than a fifty-stock simulation. A two-year simulation will probably be less correlative than an eight-year simulation. Alpha will probably be more correlative to forward returns than the information ratio. And so on. The best backtesting procedure will be the one that produces the most correlative results.
6. Because stock-market returns have never been in the least predictable, strategy returns are even less predictable. We should never try to estimate a strategy’s return. What we can do, though, is make a reasonable claim that, based on past returns, correlation studies, and robust simulations tested across different data sets, certain strategies are more likely to produce a profit than others. We should be able to assign a level of confidence to the probability that one strategy will outperform another. And we should invest in the strategy or strategies that have the best chance of outperforming.

My ten largest holdings right now: BGFV, CTG, BASI, MEIP, ODC, TATT, EGY, ELMD, TGB, FNJN.

CAGR since 1/1/2016: 56%. One-year return: 75%.