Cool, you’re definitely going on my watchlist.
Wasn’t 3xGTAA risky/volatile enough for you? xD
With two highly leveraged assets that are rebalanced once a month, there’s definitely a significant risk of drawdown.
Maybe that wasn’t the case in the past because there was always another asset with more momentum, but who knows if those correlations will hold up in the future.
What will you do if 3xQQQ and 5TLT take off and the U.S. debt crisis gets serious next month?
Have you looked at the timing of the 2000 crash or the COVID crisis? In my backtests, a strategy like that was more about luck than an edge.
Wasn’t 3xGTAA risky/volatile enough for you? xD
With two highly leveraged assets that are rebalanced once a month, there’s definitely a significant risk of drawdown.
Maybe that wasn’t the case in the past because there was always another asset with more momentum, but who knows if those correlations will hold up in the future.
What will you do if 3xQQQ and 5TLT take off and the U.S. debt crisis gets serious next month?
Have you looked at the timing of the 2000 crash or the COVID crisis? In my backtests, a strategy like that was more about luck than an edge.
•
11
•
@SemiGrowth
With MATT, you also put together the portfolio based on historical results and then optimized the parameters. How do you personally distinguish between luck and an edge?
With MATT, you also put together the portfolio based on historical results and then optimized the parameters. How do you personally distinguish between luck and an edge?
••
@Yoshika That's not what I mean. levGTAA probably has an edge. But the question is just how high the risk is. Let's say there's a coin toss. You either win 1 million or lose 900K, but you can only toss the coin once. Would you do it? You have an edge, but you can’t play it out statistically because a single negative event could wipe you out. With Matt, the risk is significantly lower, and trades are executed every 14 days.
Of course, this results in a lower return, but it significantly reduces tail risks.
In my opinion, lecGTAA has too high tail risks, but maybe you’ll get lucky.
Of course, this results in a lower return, but it significantly reduces tail risks.
In my opinion, lecGTAA has too high tail risks, but maybe you’ll get lucky.
••
@SemiGrowth
But in the case of a true overnight Black Swan event, no time interval matters—whether it’s 14 days or a month. In the end, the only thing that helps is position size. And the fact that MATT performs better in this scenario is only demonstrated by the in-sample backtest.
But in the case of a true overnight Black Swan event, no time interval matters—whether it’s 14 days or a month. In the end, the only thing that helps is position size. And the fact that MATT performs better in this scenario is only demonstrated by the in-sample backtest.
••
@Yoshika position size, leverage, and diversification. Apart from that, there was also out-of-sample data.
It starts with the undiversified universe of levGTAA
It starts with the undiversified universe of levGTAA
••
@SemiGrowth
Having four asset classes doesn't automatically mean the portfolio is undiversified. I find your comment about the out-of-sample data more interesting. Which data points were actually still untouched after MATT's entire optimization process?
Having four asset classes doesn't automatically mean the portfolio is undiversified. I find your comment about the out-of-sample data more interesting. Which data points were actually still untouched after MATT's entire optimization process?
••
@Yoshika One-tenth of the data—I’m not exactly sure which part—but I did run the out-of-sample Monte Carlo test.
levGTAA does have a significant U.S. exposure. It would be interesting to know which assets contributed to which performance. Probably 3xQQQ and BTC accounted for almost everything. The assets need to continue performing as they have over the past 20 years.
levGTAA does have a significant U.S. exposure. It would be interesting to know which assets contributed to which performance. Probably 3xQQQ and BTC accounted for almost everything. The assets need to continue performing as they have over the past 20 years.
••
@SemiGrowth
I agree with you about the U.S. risk.
But a one-tenth holdout—where you no longer know which one it was—is kind of hard to classify as "out-of-sample." And running Monte Carlo simulations on the same historical data is, to me, more of a robustness test.
I agree with you about the U.S. risk.
But a one-tenth holdout—where you no longer know which one it was—is kind of hard to classify as "out-of-sample." And running Monte Carlo simulations on the same historical data is, to me, more of a robustness test.
••
I'm out—I still need to explain to my backtest why it's not hitting 28%... Thanks, guys ʕ•ᴥ•ʔっ♡. I'm going to light another candle for the God of Overfitting 🕯️
•
22
•
@SemiGrowth @yoshika Thanks for your discussion and your questions. Yoshika, please light a candle for me too :D I tried to make a little sense of your questions for myself and answer them all at once, so I just fed the questions to the AI, and this is what came out (I actually already knew the gist of almost all the answers beforehand, but obviously I don’t know every metric by heart, so this was easier): (Backtest 1987–2026, average across all 21 possible trading days):
1) Tail risk— @SemiGrowth is right here; it’s high:
- maxDD −62% (−70% only on the last day of the month), worst month −29% (Aug. 1990), worst day −17%
- Probability of a >50% drawdown in 10 years: approx. 30%; >75%: approx. 1–2%; >90%: practically 0
While leveraged ETPs do not result in liquidation like a coin flip (no margin call; a knock-out would theoretically only be possible at −33% of the underlying asset’s value in a single day, or at 5x −20%), but you must be able to withstand losses of −60% to −80%. For me, the solution lies in position sizing and diversification across strategies.
2) Crash Timings (range across all 21 possible trading days):
- Dotcom 03/2000–10/2002: +22% (+2% to +32%); Nasdaq 100 itself −83%
- Financial Crisis 10/2007–03/2009: −4% (−23% to +14%)
- COVID 02–03/2020: −25% (−49% to −3%)
- Oct. 1987 crash: −27%, 1994 bond crash: −25%, 2022 interest rate shock: −26% (3x Nasdaq: −79%)
→ The trend filter reliably captures slow bear markets, regardless of which day trading takes place. In the case of rapid V-shaped crashes like the COVID-19 crisis, however, timing is key—and in such cases, it really is largely a matter of luck.
3) Debt crisis / Correlations: QQQ and TLT were held simultaneously on 41% of the days. Their correlation was clearly negative from 2000 to 2019 (approx. −0.3) and has been slightly positive on average since 2020 (approx. +0.2). The next historical examples of “stocks and bonds falling together” are 1994 and 2022, each with a decline of approximately −25%. A month with a −30% decline is therefore realistic. This is another reason why I’m currently allocating 3x IEF instead of 5x TLT for bonds—and also because of the costs associated with 5x TLT. The question that needs to be clarified is whether to continue using TLT as the signal provider or switch to IEF as the signal, since IEF is also the index in which we’re invested. In the future, a shift to 3x TLT—assuming it remains tradable—would be an option. Since 2020, BTC has correlated with QQQ at an average of 0.39 over a one-year period (2010–19: 0.12). The peak was 0.69 in September 2021. Since 2016, it has been above 0.5 in 24% of the months.
4) @Yoshika Contribution per asset (return contribution p.a.): Nasdaq 12.8 percentage points, TLT 7.8, BTC 7.1, gold 4.2. Nasdaq and BTC thus account for a good 60%—though this varies greatly depending on the decade: From 2000 to 2009, the majority came from gold and bonds (Nasdaq only 3.7); since 2010, it has come primarily from BTC and Nasdaq. Without the Nasdaq, the figure would be 16.4% p.a.; without BTC, 20.5%; and without TLT, 22.3%.
5) Out-of-sample: @Yoshika is right here. With levGTAA, there is also no clean out-of-sample (OOS) test; all data from 1987–2026 has been incorporated into the model, and Monte Carlo simulations on the same historical data serve as a robustness test, not an OOS test. I try to keep overfitting to a minimum (averaging parameters instead of optimizing them, plateaus, halving test, walk-forward, stress tests: BTC without trend return 23.6% p.a., bonds without interest rate tailwind 23.7%).
My comment on the AI’s response: 14-day signals were a disaster for my system, but I can’t speak for your system—it involves too much back-and-forth between assets because it switches too quickly, at the same time, it’s also a risk to include multiple highly correlated assets, since that makes back-and-forth movements between them more likely and they share a combined drawdown risk. My system, on the other hand, performs weaker when an asset underperforms over a long period. In short: no matter what you do, every system has strengths and weaknesses; I’ve decided on this approach.
This is not investment advice. It is not an advertisement for a financial product, nor is it a solicitation to buy or sell securities or certificates.
1) Tail risk— @SemiGrowth is right here; it’s high:
- maxDD −62% (−70% only on the last day of the month), worst month −29% (Aug. 1990), worst day −17%
- Probability of a >50% drawdown in 10 years: approx. 30%; >75%: approx. 1–2%; >90%: practically 0
While leveraged ETPs do not result in liquidation like a coin flip (no margin call; a knock-out would theoretically only be possible at −33% of the underlying asset’s value in a single day, or at 5x −20%), but you must be able to withstand losses of −60% to −80%. For me, the solution lies in position sizing and diversification across strategies.
2) Crash Timings (range across all 21 possible trading days):
- Dotcom 03/2000–10/2002: +22% (+2% to +32%); Nasdaq 100 itself −83%
- Financial Crisis 10/2007–03/2009: −4% (−23% to +14%)
- COVID 02–03/2020: −25% (−49% to −3%)
- Oct. 1987 crash: −27%, 1994 bond crash: −25%, 2022 interest rate shock: −26% (3x Nasdaq: −79%)
→ The trend filter reliably captures slow bear markets, regardless of which day trading takes place. In the case of rapid V-shaped crashes like the COVID-19 crisis, however, timing is key—and in such cases, it really is largely a matter of luck.
3) Debt crisis / Correlations: QQQ and TLT were held simultaneously on 41% of the days. Their correlation was clearly negative from 2000 to 2019 (approx. −0.3) and has been slightly positive on average since 2020 (approx. +0.2). The next historical examples of “stocks and bonds falling together” are 1994 and 2022, each with a decline of approximately −25%. A month with a −30% decline is therefore realistic. This is another reason why I’m currently allocating 3x IEF instead of 5x TLT for bonds—and also because of the costs associated with 5x TLT. The question that needs to be clarified is whether to continue using TLT as the signal provider or switch to IEF as the signal, since IEF is also the index in which we’re invested. In the future, a shift to 3x TLT—assuming it remains tradable—would be an option. Since 2020, BTC has correlated with QQQ at an average of 0.39 over a one-year period (2010–19: 0.12). The peak was 0.69 in September 2021. Since 2016, it has been above 0.5 in 24% of the months.
4) @Yoshika Contribution per asset (return contribution p.a.): Nasdaq 12.8 percentage points, TLT 7.8, BTC 7.1, gold 4.2. Nasdaq and BTC thus account for a good 60%—though this varies greatly depending on the decade: From 2000 to 2009, the majority came from gold and bonds (Nasdaq only 3.7); since 2010, it has come primarily from BTC and Nasdaq. Without the Nasdaq, the figure would be 16.4% p.a.; without BTC, 20.5%; and without TLT, 22.3%.
5) Out-of-sample: @Yoshika is right here. With levGTAA, there is also no clean out-of-sample (OOS) test; all data from 1987–2026 has been incorporated into the model, and Monte Carlo simulations on the same historical data serve as a robustness test, not an OOS test. I try to keep overfitting to a minimum (averaging parameters instead of optimizing them, plateaus, halving test, walk-forward, stress tests: BTC without trend return 23.6% p.a., bonds without interest rate tailwind 23.7%).
My comment on the AI’s response: 14-day signals were a disaster for my system, but I can’t speak for your system—it involves too much back-and-forth between assets because it switches too quickly, at the same time, it’s also a risk to include multiple highly correlated assets, since that makes back-and-forth movements between them more likely and they share a combined drawdown risk. My system, on the other hand, performs weaker when an asset underperforms over a long period. In short: no matter what you do, every system has strengths and weaknesses; I’ve decided on this approach.
This is not investment advice. It is not an advertisement for a financial product, nor is it a solicitation to buy or sell securities or certificates.
•
11
•