The rapid market swoon in 2020 as a result of the worldwide corona virus pandemic makes one wonder if there is any value in ever exiting the market entirely and trying to re-enter at a lower price. Statistically, the most probable outcome of such a strategy is not a good one, it is rarely beneficial. Better to let diversification take the brunt of the hit. But rarely doesn’t mean never. After building a number of competing learning machines to find optimal asset allocation over the long term, it occurred to me to build one more series of engines to explore that question; the stop loss engine. Below is a chart of the results from the machine learning engines for 2 years including the pandemic drop and recovery (2019-2021), normalized to 2019 starting values.
The learning machines spend more time in asset allocation profiles that are more aggressive than the more diversified allocations of the benchmark. As a result, when a market dislocation occurs you may see the draw down is greater with the more aggressive allocation at that moment as can be seen above, with the green (4 engine learning machine) drop being greater than the benchmark (purple). The one layer stop loss engine protected the draw-down during the pandemic induced drop, and got back invested to generate a net positive outcome.
The number of times the portfolio can dip below a recent peak is worth knowing to emotionally prepare for the inevitable drop that will always occur. Below is a plot that shows how often (occurrences per year) on average the diversified portfolio will decline a specific percentage below the previous peak value. For example, one can expect a drop of 5% from the highest peak twice per year on average.
Below is a chart comparing all the results of all the different engines between 1997 and the end of 2020.
These results include the period of the market correction that occurred during the pandemic of 2020, but the training window for the machines ended 2 years prior. The stop-loss engine did reduce the worst draw-down and improved returns, but it took a brutally fast correction and stunning rebound in the market. Although the training window ended prior to the pandemic drop, the one layer stop loss engine did generate a significant benefit at that time. Ironically, it generated no long term benefit prior to that event. The two layer stop loss engine did generate a net positive benefit throughout the training window and post training window, albeit with a signal frequency of every month or two, which may not be practical for the average long term investor.
The cells highlighted in color are green for the best result, yellow for the second best and red for the worst. The 4 engine learning machine with a 2 layer stop loss did produce the best set of overall results, although it generated the most signals on average. As a relatively conservative long term active investor, my own preference is the second best performer, which is the 4 engine learning machine with 1 layer stop loss. However, for the more active long term investor the 4 engine learning machine with a 2 layer stop loss produced the most winning set of outcomes. The stop loss engine uses machine learning to find sell and buy back levels that produce the most uniform return results. The technique becomes less and less helpful as the volatility of the target portfolio increases. This can be seen by looking at how little it helped with the more volatile benchmark portfolio.
The only difference between the 4 engine learning machine and the 5 engine machine is the latter includes a single high speed machine which generates more signals (one signal every 3 to 4 weeks on average, or one round buy/sell trip every 6 to 8 weeks). The original learning machines explored multiple speeds as a parameter and settled on a slower set of parameters as they proved to have the best performing characteristics with the most uniform returns over time. As you can see in the table the 4 engine learning machine with the 2 layer loss limit engine did improve long term results and reduced draw down, but generated a lot of trading signals not suitable for most investors. As a matter of completeness I test additional layers and reached a point of no further improvement at 4 layers.
The one layer stop loss engine generated only 2 round trips in more than 20 years. The first was in the late 90's and didn't generate any net gain. The most recent signal was during the pandemic induced market collapse in 2020, which was so fast and deep the loss limit engine did generate a net gain. But that was a very rare event and it may be another decade or two before, or if ever, we get a market drop and rebound like that.
Комментарии