Last time I set the stage for “Actuarial Harvesting” — an attempt to improve on McClung’s Prime Harvesting but without the memory effect. The preliminary results didn’t look good, though.
It came in dead last.
After some digging, I discovered I had made two mistakes:
- There was a mistake in the Actuarial Harvesting algorithm that was causing it to harvest “too much”. This had the result of keeping the portfolio extremely heavy on bonds. Usually 70–80% bonds. Well, crap. So I fixed that.
- I also realised there was a problem with my Monte Carlo simulation. Previously I had used the same code to test different withdrawal strategies. When you do that, the asset allocation stays the same. But when you’re testing harvesting strategies, the asset allocation changes. After all, that’s the whole point of a harvesting strategy. But the Monte Carlo simulation wasn’t changing the asset allocation properly. Well, crap again. So I fixed that as well.
Where does that leave us?
Bootstrapping
For the first test, we’ll use historical bootstrapping. That is, we’ll take actual historical returns but randomly shuffle them up. This allows to create thousands of variations and test them.
And we’ll test it against US, UK, and Japan data sets. This test uses variable withdrawals and the metric is HREFF-4, which measures how often our withdrawals fall below 4%; a higher score is better.
It turns out that Actuarial Harvesting does pretty well overall — much better than my previous post suggested. That said, notice that most of the strategies have pretty similar results. In the charts above there are some strategies that are clearly “not good” but everything else seems to be fine.
But there’s something else a bit fishy about the above charts. Notice that 100% stocks always comes in first? And notice that “110-age in bonds” always beats out “100-age in bonds”? And “120-age in bonds” always beats “110-age in bonds”?
It sure looks like the most important thing is just how high your stock percentage is, rather than any fancy harvesting.
Historical Data
The test above took historical data and shuffled up the order. So sometimes 1952 might come after 2003. Let’s try the test again but preserve the actual historical order.
For the US there’s no real change in the ordering. The 4 or 5 “not good” strategies look even worse on this test.
The same is true for the UK. The ordering is essentially the same and the “not good” strategies lose big a bigger gap.
For Japan, though, things look pretty different. Prime Harvesting is substantially better than anything else and the strategies that are more equity-heavy seem fare poorly. We see two things happen:
- When using the bootstrapped simulation, Japan had similar results to the US and UK. But when we switch to historical ordering, Japan looks quite different from the US and UK.
- When looking at the US and the UK they have the same results for bootstrapping versus historical ordering. But Japan looks different from itself: the bootstrapping and historical ordering give us fairly different results.
One plausible explanation for this gap is that Japan has been historically “unlucky” — so the exact sequence of returns has an outsize impact. Whereas the US and UK have been “normal” so random ordering doesn’t change things much.
These results — using HREFF-4 — are similar to what I found in an earlier investigation of harvesting strategies.
It looks like Actuarial Harvesting isn’t as bad as I initially thought. But it also isn’t some silver bullet. It does okay but hardly stands out from the crowd, despite being theoretically more plausible than, say, Prime Harvesting. Overall, it is hard not to feel like none of these harvesting strategies really add much value, though. The top contenders all seem to be nearly indistinguishable by the metrics we’ve looked at. Is there some other way of judging success?
But there’s another set of concerns — Actuarial Harvesting (at least the way we’ve defined it so far) — generally leaves you with a very small bond allocation. On average, you are 87% stocks and 13% bonds (87/13). Other strategies tend to keep you with a higher bond allocation:
- Prime Harvesting’s average asset allocation is 71/29
- OmegaNot’s average asset allocation is 76/24
- Weiss’s average asset allocation is 70/30
We can tweak that: we’re currently using a discount rate of 5.4% in our Actuarial Harvesting calculations. There’s nothing magic about that. There was some reasoning behind it but nothing says it is “right”. If we use a higher discount rate, say 9%, then we end up with more bonds on average (65/35)…but then it doesn’t rate as highly on the HREFF-4 metric.
It seems clear that there is an almost direct relationship between how few bonds you hold and how well the retirement fares on a wide variety of metrics.
When we look at 100% stocks it is:
- #1 on HREFF-4 for US bootstrapping
- #1 on HREFF-4 for UK bootstrapping
- #1 on HREFF-4 for Japan bootstrapping
- #1 on HREFF-4 for US historical
- #1 on HREFF-4 for UK historical
And from the other harvesting testing I did it is also…
- #1 on US mean & median income
- #1 on UK mean & median income
- #1 on UK 5th and 10th percentile income (i.e. the worst case scenarios)
- #1 on Japan mean & median income
The interesting question is why an all‐equity portfolio is considered such a risky alternative. The answer, in fact, is far from clear.
In “The Retirement Glidepath”, Javier Estrada looks at similar data (though he looks at many countries, not just these three) and comes to a similar conclusion. In the section “Why Not 100% Stocks?” he suggests that “having a retirement portfolio fully invested in stocks is a strategy that should be seriously considered by retirees”.
It is a position that I’ve become more and more sympathetic towards.
Should we all just be 100% stocks? Are there some bad scenarios or “super worst case scenarios” where being 100% stocks really hurt? Maybe one or two of those super black swans are what we’re protecting ourselves against and those black swans get swamped in “normal” data when we look at bigger data sets.
100% stocks does pretty well on the 10th percentile income and even the 5th percentile income. But maybe it falls apart on 1st percentile or 0.1st percentile income or during 1966 or some other “once in a thousand years” type event.
