1Yr
Excellent post, I'm super interested in doing some research into this
•
11
•
1Yr
@Epi Something I'm considering is having many unrelated asset classes (10+), then for each asset class with positive relative (3 months) and absolute (12 months) momentums, we take product of both the momentums, compare with the product of other positive asset classes, then assign a percentage proportional to the sum of the products.
••
1Yr
@000 Sounds a bit complicated. What is the idea behind the calculation? Normally the performances of the time periods are simply added up, if necessary normalized before. As already said in the article, I like it simple. Since GTAA is not easy to understand in detail anyway, I would not add unnecessary complexity at this point. My goal is a model that trades common ETFs 1x per month and makes 20%pa at 10-15% max drawdown in backtest since 2001. I'm close, but haven't made it yet (19.1% at 19% maxDD). How far does your proposal get in the backtest?
••
1Yr
@000 Your assumption of choosing 10 uncorrelated asset classes sounds good. And I would be very grateful if you could name them briefly. I have searched and tried a lot, but there are just not many uncorrelated classes. Stocks among themselves are correlated like 0.6-1, Treasuries depending on maturity -0.2-0.4, Commodities, Corp Bonds, Reits and Crypto are more correlated with stocks. The only asset that is uncorrelated to almost everything else is gold. Finding 10 uncorrelated asset classes here is really not easy!
••
1Yr
@Epi I agree, my proposal is just a hypothetical currently, however I'm interested in learning more about your approach. Are you investing in the 3 ETFs you used for your backtest? I am curious to learn more about why you chose LSGBX and ^GOLD specifically. The issue with backtesting is that past performance doesn't dictate future performance, and there are many combinations that outperform SPY, but are not as diversified. Additionally, different time frames for momentum and proportions significantly changes performance, which makes the strategy not very robust, in my opinion.
•
11
•1Yr
@000 Thanks for the suggestions! I am investing in a different allocation. The example was just for illustration and a starting point for my own research. why LSGBX and gold? Both are highly correlated with world bonds and commodities respectively (around 0.9 I think) and the data series is long enough for a decent backtest. Instead of SPY, I would have preferred an AllWorldETF, but they only go back to 2005 or so. I would like to see how a model performs in the 2000s and 2008 crash, though. You raise an important and difficult point with robustness. There are some high performing models, but they crash dramatically with the smallest variations in the time series. These fall out for me. I want a model that can handle smaller variations well. Example: the model mentioned in the post is not very robust over the momentum periods. This is usually the case if you only take a few periods. On the other hand, the combos 1-3-6-12months and from top 2 are quite stable. To discard the strategy as a whole as not very robust if one model interpretation is unstable, I think is premature. You have to do some searching and research. Also in terms of content - i.e. why some combos work well and some don't - and then find a model interpretation. Difficult but exciting.
••
1Yr
@Epi I greatly appreciate your insight. I've read your other comments on this post and learned a lot. One thing I'm struggling to comprehend is allocating portfolio 100% of asset class with best sum of momentums, instead of proportionally allocating based on sum. Can you elaborate on the logic behind this?
•
11
•1Yr
@000 Hmm, I think the idea is simple: always invest in the asset with the strongest positive momentum. You can also take the top 2 or 3, which has a positive effect on the Sharpe ratio with more asset classes (>4). I find your idea of making the allocation proportional to the momentum totals very interesting. Consequently, you would also have to short if the momentum total is negative - which unfortunately (or fortunately) is not easy in Germany. Indirectly one could do this by including ShortEFs in the allocation, but my models have never done better with these, rather they have collapsed dramatically. Otherwise, your idea could create a whole new timing model, namely if the asset allocation constantly adjusts, not just once a month, but daily or weekly. But that would have to be done by a robot, I don't have that much time. At the moment, the trading costs are probably still too high. Maybe this is the future of investing, but your idea should be tested. Intuitively I find it convincing. Do you have an idea how this could be done with manageable effort?
••