Where Risk Parity Hurts: A 58-Year Audit of Tails and Drawdowns
The previous article extended the inverse-volatility allocation across SPY, TLT, and GLD back to 1968 using a synthetic price construction. Over fifty-eight years the strategy delivered a CAGR of 7.1%, volatility of 7.5%, a Sharpe of 0.97, and a maximum drawdown of 22%. The volatility-targeting overlay, justified by the persistence of volatility across the same window, kept realised vol close to the 8% target through regimes that look very different from each other.
The drawdown record across this window is the object of this article. The 22% maximum sits at the worst point of one episode, but it is one number, taken from one moment, in one configuration of assets — and notably, the worst drawdown in the record arrives before gold was a tradable holding. Reading the audit through any single number flattens what is actually there. The worst regions of the distribution are populated by drawdowns of different magnitudes, different durations, and different asset-level signatures, and the audit becomes informative once each is read on its own terms. What kind of months produced the worst monthly returns. What kind of days produced the worst daily returns. What the assets were doing when each large drawdown unfolded.
The reason to do the audit before adding anything is concrete. Each historical drawdown is a sample of a regime the strategy will encounter again. Reading those samples at the asset level produces a catalogue of failure modes that any candidate addition can be tested against — an addition that does not produce a useful return inside one of these regimes does not address the part of the distribution where the core suffers. Building sleeves with that catalogue as a target is also the addition I would feel comfortable holding through the next twenty-percent drawdown.
Two grids of pain
To read structure beyond the drawdown number we look at the return distribution itself and ask where the worst observations sit, and what they have in common. We do this on two grids. The daily grid catches acute single-day events — flash crashes, FOMC surprises, geopolitical shocks. The monthly grid catches sustained regime changes — multi-month corrections where each individual day looks survivable but the aggregate becomes destructive.
For each grid we look at CVaR(5%), the average return on the worst five percent of observations. Each tail observation is marked on the equity curve and classified by which assets drove the loss.
The daily picture
The daily scatter plot separates cleanly into regions. All-down days cluster in the lower-left where both stocks and bonds are negative; bonds-and-gold-down days fall in the lower-right where bonds drop with stocks rising; stocks-and-gold-down days appear in the upper-left where bonds rally as a haven. The boundaries between regimes are visible objects, not artefacts of classification.
Each region has a different defensive logic. An all-down day rewards being out of all three assets. A bonds-and-gold-down day with stocks rising rewards staying in stocks. A stocks-and-gold-down day with bonds rallying rewards holding bonds. The strategy as built does not distinguish — the volatility scalar treats all of these stresses the same way.
The monthly picture
The monthly grid is structurally distinct from the daily grid. Daily tails are largely background noise — they happen with predictable frequency throughout the timeline. Monthly tails are concentrated regime breaks. They occur when something in the macro environment changes and the change propagates across asset classes for weeks or months at a time.
The shift in regime composition between the two grids carries the analytical content. The fact that stocks-and-bonds-together rises from 14% at the daily horizon to 23% at the monthly horizon means the 2022 inflation shock pattern — stocks and bonds both falling for an extended period as the macro regime changes — is not just a daily phenomenon that integrates up. It has its own monthly identity, with its own historical instances in 1973–74, 1980, and at points throughout the early 1990s.
Six worst drawdowns
The CVaR figures show distributional structure but compress the timeline. To see what each tail event actually was, we look at the six largest peak-to-trough drawdowns in the strategy’s history and decompose them by asset.
None of the six is a repeat of any other. The two all-down episodes differ in time scale — 1968–70 was a slow twenty-month grind, 2022 was a rapid ten-month duration crash. The two stocks-and-bonds-together episodes differ in gold behaviour — gold surged in 1972–74, gold was flat in 1987. The 1983–84 bonds-and-gold mode appears nowhere else in the record. The 2008 equities-and-gold episode with bonds rallying as flight-to-quality has no other instance in the table.
The diversity matters more than the magnitudes. A reader looking at the six bars together can see that the strategy has lost similar amounts in regimes that have nothing else in common. The asset-level decomposition turns each bar into a story about what the macro environment was doing to which exposure, and over what time scale.
What the audit gives us
The audit produces two regimes worth naming explicitly: the daily 5% CVaR regime, the monthly 5% CVaR regime. Each has its own asset-level signature, its own time scale, and its own historical instances. Future work on this allocation can use those two regimes as design coordinates — return streams that operate inside one or both of them, with payoffs that complement what the core delivers there.
Two existing pieces of work already touch this space. The Turnaround Tuesday strategy deploys capital primarily inside high-stress windows — its activity overlaps with the daily CVaR regime by construction, even though it was not designed against the multi-asset core specifically. The short-TLT leg of the turn-of-month strategy carries a structurally negative bond exposure into the regimes where bond losses drive monthly drawdowns; that overlap is incidental rather than targeted, but the architectural pattern is the same one the monthly regime invites.
Neither sleeve closes the problem. Both demonstrate that designing against a specific tail regime, rather than against the strategy in aggregate, is a workable mode of construction. The work going forward is to engineer the return distribution around this core — adding return streams whose payoffs land inside the daily and monthly CVaR regimes documented above, and stress-testing every candidate addition against the same catalogue of historical failures the audit has produced.