This point is especially obvious if one looks at some of the results from here, where I looked at different weights with monthly rebalancing. Some of the other posts look at other weighting schemes.langlands wrote: ↑Sun Aug 09, 2020 2:57 pmMuch much more likely that it's because of backtesting overfit. If end of quarter actions could have such a large (and reliable) effect, the markets would be ridiculously inefficient and we should all be doing market timing like clock-work.
Sure, it's debatable whether this is a real effect. But I would stick to it going forward regardless, out of superstition if nothing else, if I were doing quarterly rebalancing. It shouldn't hurt, at least.
Others have noticed the monthly cyclicity in bonds and equities. I would not be at all surprised if this effect may already be disappearing or have disappeared already.
The fact that different dates of rebalancing can have such a big impact on the strategy (as noticed also by many participants of this strategy who've traded it live over the past year) shows that backtesting a single "representative" of the strategy from the class of all equivalent such strategies is extremely inadequate. Basically, the variance in outcomes is massive and more independent Monte Carlo simulations are needed to get any semblance of statistical significance. Comparing single backtest results between arbitrary instantiations of different strategies is meaningless because there is probably just as much variation between instantiations of the same strategy as there is between strategies.
The figure in that post shows every 5-year sequence with 21-day rebalancing over the history of the simulated daily UPRO & TMF (starting end of 1986), looking at combinations from 30/70 to 60/40 UPROSIM/TMFSIM compared to the original 40/60 UPROSIM/TMFSIM with quarterly rebalancing.
I'd post the image again here but it seems to be too large.
The second plot on the right in the linked figure shows the cumulative distribution of all of the 5-year sequences. The squares are the mean of the distribution and the vertical bars on the bottom show the median for the distributions. The quarterly rebalance set is right in the middle of the weighted sets. The key is that the spread in 5-year outcomes is way way larger than the spread in the different weighting schemes, which implies that the time period is way way more important than the weighting scheme. In other words, "the variance in outcomes is massive."
The weighting in UPRO versus TMF didn't greatly affect the mean/median CAGR (except that the blend performed better than either UPRO or TMF in isolation), but did substantially affect the spread in 5-year CAGR. The 30/70 scheme had almost all outcomes between 0 and 30 CAGR, the 60/40 had almost all outcomes between -10 and 50 CAGR. In other words, a greater weight in UPRO gave a small increase in mean/median CAGR with a substantial increase in the spread of results. I'd expect that the effect of weighting on the spread will probably be fairly robust going forward, although the mean/median value is less robust.